Image Analysis and Processing – ICIAP 2005

Structural pattern representations, especially graphs, have advantages over feature vectors. However, they also suffer from a number of disadvantages, for example, their high computational complexity. Moreover, we observe that in the field of statistical pattern recognition a number of powerful concepts emerged recently that have no equivalent counterpart in the domain of structural pattern recognition yet. Examples include multiple classifier systems and kernel methods. In this paper, we survey a number of recent developments that may be suitable to overcome some of the current limitations of graph based representations in pattern recognition.

Business development in highly competitive new markets is strongly dependent on the level of technology innovation. Actually, the real success of a new product is affected by many other factors in the development chain, including customer requirements understanding and system implementation. The paper is aimed to refer some direct experience of New product development based on Computer Vision Technology. A practical example is referred to describe a recent success story in the field of Security application, a mobile Automatic Number Plate Recognition system that is currently used by many security forces in different countries in Europe and the USA.

Reliable person recognition is an integral component of identity management systems. Biometrics offers a natural and reliable solution to the problem of identity determination by recognizing individuals based on their physiological and/or behavioral characteristics that are inherent to the person. Although biometric systems have been successfully deployed in a number of civilian applications, current biometric systems are not perfect. In this paper, we describe the various obstacles that prevent biometric systems from achieving foolproof automatic person recognition. We also show that using multiple biometric modalities can alleviate some of the problems faced by unimodal biometric systems. Finally, we present the vulnerabilities of biometric systems and discuss solutions to protect biometric systems from some common attacks.

In conventional computer vision systems symbol grounding is invariably established via supervised learning. We investigate unsupervised symbol grounding mechanisms that rely on perception action coupling. The mechanisms involve unsupervised clustering of observed actions and percepts. Their association gives rise to behaviours that emulate human action. The capability of the system is demonstrated on the problem of mimicking shape puzzle solving. It is argued that the same mechanisms support unsupervised cognitive bootstrapping in cognitive vision.

As the accuracy of conventional classifiers, based only on a static partitioning of feature space, appears to be approaching a limit, it may be useful to consider alternative approaches. Interactive classification is often more accurate then algorithmic classification, and requires less time than the unaided human. It is more suitable for the recognition of natural patterns in a narrow domain like trees, weeds or faces than for symbolic patterns like letters and phonemes. On the other hand, symbolic patterns lend themselves better to using context and style to recognize entire fields instead of individual patterns. Algorithmic learning and adaptation is facilitated by accurate statistics gleaned from large samples in the case of symbolic patterns, and by skilled human judgment in the case of natural patterns. Recent technological advances like pocket computers, camera phones and wireless networks will have greater influence on mobile, distributed, interactive recognition of natural patterns than on conventional high-volume applications like mail sorting , check reading or forms processing.

Application and development of specialized machine learning techniques is gaining increasing attention in the intrusion detection community. A variety of learning techniques proposed for different intrusion detection problems can be roughly classified into two broad categories: supervised (classification) and unsupervised (anomaly detection and clustering). In this contribution we develop an experimental framework for comparative analysis of both kinds of learning techniques. In our framework we cast unsupervised techniques into a special case of classification, for which training and model selection can be performed by means of ROC analysis. We then investigate both kinds of learning techniques with respect to their detection accuracy and ability to detect unknown attacks.

Intrusion Detection Systems (IDSs) play an essential role in today’s network security infrastructures. Their main aim is in finding out traces of intrusion attempts alerting the network administrator as soon as possible, so that she can take suitable countermeasures. In this paper we propose a

misuse-based

Network Intrusion Detection architecture in which we combine multiple one-class classifiers. Each one-class classifier is trained in order to discriminate between a specific attack and all other traffic patterns. As attacks can be grouped in classes according to a taxonomy, for each attack class a number of one-class classifiers are trained, each one specialized to a specific attack. The proposed multiple classifier architecture combine the outputs of one class classifiers to attain an IDS based on

generalized attack signatures

. The aim is in labelling a pattern either as normal or as belonging to one of the attack classes according to the adopted taxonomy. The potentials and effectiveness of the proposed approach are analysed and discussed.

When addressing the problem of detecting malicious activities within network traffic, one of the main concerns is the reliability of the packet classification. Furthermore, a system able to detect the so-called

zero-day attacks

is desirable. Pattern recognition techniques have proven their generalization ability in detecting intrusions, and systems based on multiple classifiers can enforce the detection reliability by combining and correlating the results obtained by different classifiers.

In this paper we present a system exploiting genetic algorithms for deploying both a misuse-based and an anomaly-based classifier. Hence, by suitably combining the results obtained by means of such techniques, we aim at attaining a highly reliable classification system, still with a significant degree of new attack prediction ability. In order to improve classification reliability, we introduce the concept of rejection: instead of emitting an unreliable verdict, an ambiguous packet can be logged for further analysis. Tests of the proposed system on a standard database for benchmarking intrusion detection systems are also reported.

EnFilter is a Proactive Password Checking System, designed to avoid password guessing attacks. It is made of a set of configurable filters, each one based on a specific pattern recognition measure that can be tuned by the system administrator depending on the adopted password policy. Filters use decision trees, lexical analysers, as well as Levenshtein distance based techniques. EnFilter is implemented for Windows 2000/2003/XP.

The continuous evolution of the attacks against computer networks has given renewed strength to research on anomaly based Intrusion Detection Systems, capable of automatically detecting anomalous deviations in the behavior of a computer system. While data mining and learning techniques have been successfully applied in host-based intrusion detection, network-based applications are more difficult, for a variety of reasons, the first being the curse of dimensionality. We have proposed a novel architecture which implements a network-based anomaly detection system using unsupervised learning algorithms. In this paper we describe how the pattern recognition features of a Self Organizing Map algorithm can be used for Intrusion Detection purposes on the payload of TCP network packets.

Security information about information systems provided by current intrusion detection systems (IDS) is spread over numerous similar and fine-grained alerts. Security operators are consequently overwhelmed by alerts whose content is too poor. Alarm correlation techniques are used to reduce the number of alerts and enhance their content. In this paper, we tackle the alert correlation problem as an information retrieval problem in order to make the handling of alert groups easier.

Fan-Meshes (FM) are a kind of geometrical primitives for generating 3D model or scene descriptions that are able to preserve both local geometrical details and topological structures. In this paper, we propose an efficient simplification algorithm for the FM models to achieve fast post-processing and rendering of large models or scenes. Given a global error tolerance for the surface approximation, the algorithm can find an approximately minimal set of FMs that covers the whole model surfaces. As compared with splat-based methods, the FM description has a large simplification rate under the same surface fitting error measurement.

The fruition of archaeological caves, hardly accessible by visitors, can benefit from a mobile vehicle which transmits to users located outside a continuous stream of images of the cave that can be visually integrated with information and data to increase the fruition and understanding of the site. This application requires self-positioning the vehicle with respect to a target. Preserving the cave imposes the use of natural landmarks as reference points, possibly using uncalibrated techniques. We have applied the modified POSIT algorithm (camera pose estimation method using uncalibrated images) to self-position the robot. To account for the difficulty of evaluating natural landmarks in the cave the tests have been made using a photograph of the prehistoric wall paintings of the archeological cave “Grotta dei Cervi”. The modified version of the POSIT has been compared with the original formulation using a suitably designed grid target. Therefore the performance of the modified POSIT has been evaluated by computing the position of the robot with respect to the target on the base of feature points automatically identified on the picture of a real painting. The results obtained using the experimental tests in our laboratory are very encouraging for the experimentation in the real environment.

In this paper a 3D graphics-based remote collaborative environment is introduced. This system is able to provide multiclient and multimedia communication, and exploits a novel multimodal user interaction paradigm based on hand gesture and perceptual user interfaces. The use of machine vision technologies and a user-centered approach produce a highly usable and natural human-computer interface, allowing even untrained users a realistic and relaxing experience for long and demanding tasks. We then note and motivate that such an application can be considered as an Augmented Reality application; according to this view, we describe our platform in terms of long-term usability and comfort of use. The proposed system is expected to be useful in remote interaction with dynamic environments. To illustrate our work, we introduce a proof-of-concept multimodal, bare-hand application and discuss its implementation and the obtained experimental results.

We present a system locating the contour of an iris in space using robust active ellipse search and correspondenceless stereo. Robust iris location is the basis for gaze estimation and tracking, and, as such, an essential module for augmented and virtual reality environments. The system implements a robust active ellipse search based on a multi-scale contour detection model. The search is carried out by a simulated annealing algorithm, guaranteeing excellent performance in spite of heavy occlusions due to blinking, uncontrolled lighting, erratic target motion, and reflections of unpredictable scene elements. Stereo correspondence is avoided altogether by intersecting conjugate epipolar lines with the located ellipses. Experiments on synthetic and real images indicate very good performance of both location and reconstruction modules.

In this paper, we propose a system for inputting hand drawn pattern and displaying it on a hand-held planar rectangle board by utilizing two cameras and a projector which are installed above the user. In this system, the cameras and the projector are related by projective geometry. The cameras capture the user’s hand-drawing motion. From the captured images, the tip of the user’s pen and the corners of the board are tracked for detecting the status of pen’s up/down and displaying the drawn pattern at a fixed position on the surface of the board. For demonstrating the validity of the proposed system, we show that the user can draw pattern on the board with arbitrary pose and position, while the pattern is simultaneously displayed on the board in real-time.

We present an uncalibrated projector-camera system in which the information displayed onto a planar screen can be interactively warped according to an arbitrary planar homography. The user interacts with the system through a laser pointer, whose displacements on the screen plane are captured by the camera, interpreted as mouse drags, and used to control the warping process. Applications of our interactive warping system encompass arbitrary (pan/tilt/screw) keystone correction, visualization of undistorted information for a user being in a general position with respect to the screen (including

virtual anamorphosis

as a special case), and self-shadow avoidance by a nearly-parallel projection.

The future customer service provided by call centres will be changed due to new web-based interactive multimedia technologies. Technical support will be offered in a completely new way by using advanced image processing technologies and natural representation of virtual humans. We present a prototype system of an animated avatar, which is steered by live body motion of the operator in a call centre. The hand and head motion is transferred directly to the avatar at the customer side in order to support a more natural representation of the virtual human. The system tracks the operators hands and the head motion quite robust in real-time without specific initialization based on a monocular camera.

We propose a novel vision-based registration approach for Augmented Reality with integration of arbitrary multiple planes. In our approach, we estimate the camera rotation and translation by an uncalibrated image sequence which includes arbitrary multiple planes. Since the geometrical relationship of those planes is unknown, for integration of them, we assign 3D coordinate system for each plane independently and construct projective 3D space defined by projective geometry of two reference images. By integration with the projective space, we can use arbitrary multiple planes, and achieve high-accurate registration for every position in the input images.

A novel algorithm, based on Kalman filtering is presented for updating the background image within video sequences. Unlike existing implementations of the Kalman filter for this task, our algorithm is able to deal with both gradual and sudden global illumination changes. The basic idea is to measure global illumination change and to use it as an external control of the filter. This allows the system to better fit the assumptions about the process to be modeled. Moreover, we propose methods to estimate measurement noise variance and to deal with the problem of saturated pixels, to improve the accuracy and robustness of the algorithm. The algorithm has been successfully tested in a traffic surveillance task by comparing it to a background updating algorithm, based on Kalman filtering, taken from literature.

We present a technique that achieves local contrast enhancement by representing it as an optimization problem. For this, we first introduce a scalar objective function that estimates the

average local contrast

of the image; to achieve the contrast enhancement, we seek to maximize this objective function subject to strict constraints on the local gradients and the color range of the image. The former constraint controls the amount of contrast enhancement achieved while the latter prevents over or under saturation of the colors as a result of the enhancement. We propose a greedy iterative algorithm, controlled by a single parameter, to solve this optimization problem. Thus, our contrast enhancement is achieved without explicitly segmenting the image either in the spatial (multi-scale) or frequency (multi-resolution) domain. We demonstrate our method on both gray and color images and compare it with other existing global and local contrast enhancement techniques.

In this paper we introduce a model-based background subtraction approach where first silhouettes, which model the correlations between neightboring pixels are being learned and where then Bayesian propagation over time is used to select the proper silhouette model and tracking parameters. Bayes propagation is attractive in our application as it allows to deal with uncertainties in the video data during tracking. We eploy a particle filter for density estimation. We have extensively tested our approach on suitable outdoor video data.

Error bounds for estimation of moments from a fuzzy representation of a shape are derived, and compared with estimations from a crisp representation. It is shown that a fuzzy membership function based on the pixel area coverage provides higher accuracy of the estimates, compared to binary Gauss digitization at the same spatial image resolution. Theoretical results are confirmed by a statistical study of disks and squares, where the moments of the shape, up to order two, are estimated from its fuzzy discrete representation. The errors of the estimates decrease both with increased size of a shape (spatial resolution) and increased membership resolution (number of available grey-levels).

Feature matching and pose estimation are two crucial tasks in computer vision. The widely adopted scheme is first find the correct matches then estimate the transformation parameters. Unfortunately, such simple scheme does not work well sometimes, because these two tasks of matching and estimation are mutually interlocked. This paper proposes a new method that is able to estimate the transformation and find the correct matches simultaneously. The above interlock is disentangled by an alternating Newton iteration method. We formulate the problem as a nearest-matrix problem, and provide a different numerical technique. Experiments on both synthetic and real images gave good results. Fast global convergence was obtained without the need of good initial guess.

Camera calibration is a necessary step in 3D modeling in order to extract metric information from images. Computed camera parameters are used in a lot of computer vision applications which involves geometric computation. These applications use camera parameters to estimate the 3D position of a feature in the image. Depending on the accuracy of the computed camera parameter, the precision of the position of the image feature in the 3D scene vary. Moreover if previously the accuracy of camera parameters is known, one technique or another can be choose in order to improve the position of the feature in the 3D scene.

Calibration process consists of a closed form solution followed by a non linear refinement. This non linear refinement gives always the best solution for a given data. For sure this solution is false since input data is corrupted with noise. Then it is more interesting to obtain an interval in which camera parameters are contained more than an accurate solution which is always false.

The aim of this paper is to present a method to compute the interval in which the camera parameter is included. Computation of this interval is based on the residual error of the optimization technique. It is know that calibration process consists of minimize an index. With the residual error of the index minimization an interval can be computed in which camera parameter is. This interval can be used as a measurement of accuracy of the calibration process.

Moment invariants are properties of connected regions in binary images that are invariant to translation, rotation and scale. They are useful because they define a simply calculated set of region properties that can be used for shape classification and part recognition. Orthogonal moment invariants allow for accurate reconstruction of the described shape. Generic Fourier Descriptors yield spectral features and have better retrieval performance due to multi-resolution analysis in both radial and circular directions of the shape. In this paper we first compare various moment-based shape description techniques then we propose a method that, after a previous image partition into classes by morphological features, associates the appropriate technique with each class, i.e. the technique that better recognizes the images of that class. The results clearly demonstrate the effectiveness of this new method regard to described techniques.

In this paper, we describe a compact circuit for real-time detection of line segments using the Line Hough Transform (LHT). The LHT is a technique to find out lines in an image. The LHT is robust to noises, but requires long computation time. The circuit calculates (1)

r

and

θ

of lines (

r

is the distance from the origin to a line and

θ

is the angle of the line) by the LHT units in parallel, and (2) start and end points of the lines by the other units which are completely pipelined with the LHT units. With this parallel and pipeline processing, the circuit can detects line segments by

π

/512 angle steps in a standard size image (640 × 480) in real-time. This circuit was implemented on an off-the-shelf PCI board with one Field Programmable Gate Array (FPGA) chip. The size of the circuit is 45% of the chip, which makes it possible to implement other circuits for higher level processing of object recognition on the same chip, or the performance can be improved twice by using the rest of hardware resources.

2D grey-level images are interpreted as 3D binary images, where the grey-level plays the role of the third coordinate. In this way, algorithms devised for 3D binary images can be used to analyse 2D grey-level images. Here, we present three such algorithms. The first algorithm smoothes a 2D grey-level image by flattening its geometrical and grey-level peaks while simultaneously filling in geometrical and grey-level valleys, regarded as non significant in the problem domain. The second algorithm computes an approximation of the convex hull of a 2D grey-level object, by building a covering polyhedron closely fitting the corresponding object in a 3D binary image. The result obtained is convex both from the geometrical and grey-level points of view. The third algorithm skeletonizes a 2D grey-level object by skeletonizing the top surface of the object in the corresponding 3D binary image.

Because of perspective distortion between the camera and the license plate, slanted images commonly appear in License Plate Recognition System (LPR system), and it seriously affects the recognition result. This paper presents two types of image slant, and gives an efficient way to rectify them. For horizontal slant correction, the proposed method is based on connected areas labeling and straight-line fitting. For vertical slant correction, the method is based on rotation experiments with various angles. Practical use in the LPR system shows that this method has a correct rate above 97%.

This paper presents an approach for speckle reduction and coherence enhancement of ultrasound images based on total variation (TV) minimization. The proposed method can preserve information associated with resolved object structures while reducing the speckle noise. However, since the equation system deduced by the TV-based method is a strongly nonlinear partial differential equation (PDE) system, the convergence rate is very slow when using standard numerical optimization techniques. So in this paper, we introduce the nonlinear multi-grid algorithm to solve this system. Numerical results indicate that the image can be recovered with satisfied result even contamination of strong noise using the proposed method and the algorithm of nonlinear multi-grid has more efficiency than the conventional numerical techniques such as conjugate gradient (CG).

In this paper a novel statistical shape feature called the Contour Co-occurrence Matrix (CCM) is proposed for image classification and retrieval. The CCM indicates the joint probability of contour directions in a chain code representation of an object’s contour. Comparisons are conducted between different versions of the CCM and several other shape descriptors from e.g. the MPEG-7 standard. Experiments are run with two defect image databases. The results show that the CCM can efficiently represent and classify the difficult, irregular shapes that different defects possess.

In this paper we concentrate on a measure of symmetry. Given a transform

S

, the kernel

SK

of a pattern is defined as the maximal included symmetric sub-set of this pattern. A first algorithm is outlined to exhibit this kernel. The maximum being taken over all directions, the problem arises to know which center to use. Then the optimal direction triggers the shift problem too. As for the measure we propose to compute a modified difference between respective surfaces of a pattern and its kernel. A series of experiments supports actual algorithm validation.

Digital cameras are gaining in popularity, and not only experts in image analysis, but also the average users, show a growing interest in image processing. Many different kinds of software for image processing offer tools for object selection, or segmentation, but most of them require expertise knowledge, or leave too little freedom in expressing the desired segmentation. This paper presents an easy to use tool for object segmentation in color images. The amount of user interaction is minimized, and no tuning parameters are needed. The method is based on the watershed segmentation algorithm, combined with seeding information given by the user, and color space projections for optimized object edge detection. The presented method can successfully segment objects in most types of color images.

We investigate the properties of edge preserving smoothing in the context of Finite Markov Random Fields (FMRF). Our main result follows from the definition of discontinuity adaptive potential for FMRF which imposes to penalize linearly image gradients. This is in agreement with the Total Variation based regularization approach to image recovery and analysis. We also report a fast computational algorithm exploiting the finiteness of the field, it uses integer arithmetic and a gradient descent updating procedure. Numerical results on real images and comparisons with anisotropic diffusion and half-quadratic regularization are reported.

In this paper, we describe a compact system for high speed computation of the optical flow. This system consists of one off-the-shelf PCI board with one Field Programmable Gate Array (FPGA) chip, and its host computer. With this system, we can generate dense vector maps at (1) 840 frames per second (fps) in small size (320 × 240) images, and (2) 30 fps in standard size (640 × 480) images by configuring different circuits on the FPGA chip. In the two circuits, vectors for all pixels in the images are obtained by the area-based matching (windows of 7 × 7 pixels are compared with 121 and 441 windows in the target image respectively). The circuits implemented on the FPGA do not require any special hardware resources, and can be implemented on many off-the-shelf FPGA boards shipped from many vendors. This system can also be used for the stereo vision by slightly modifying the circuits, and achieve the same performance.

Standard image processing algorithms for digital images require the availability of complete, and regularly sampled, image data. This means that irregular image data must undergo reconstruction to yield regular images to which the algorithms are then applied. The more successful image reconstruction techniques tend to be expensive to implement. Other simpler techniques, such as image interpolation, whilst cheaper, are usually not adequate to support subsequent reliable image processing. This paper presents a family of autonomous image processing operators constructed using the finite element framework that enable direct processing of irregular image data without the need for image reconstruction. The successful use of reduced data (as little as 10% of the original image) affords rapid, accurate, reliable, and computationally inexpensive image processing techniques.

We introduce three new texture features that are based on the morphological scale-space operator known as the sieve. The new features are tested on two databases. The first, the Outex texture database, contains Brodatz-like textures captured under constant illumination, scale and rotation. The second, the Outex natural scene database, contains images of real-world scenes taken under variable conditions. The new features are compared to univariate granulometries, with which they have some similarities, and to the dual-tree complex wavelet transform, local binary patterns and co-occurrence matrices. The features based upon the sieve are shown to have the best overall performance.

Enhancement of noisy images using a sliding discrete cosine transform (DCT) is proposed. A minimum mean-square error estimator in the domain of a sliding DCT for noise removal is derived. This estimator is based on a fast inverse sliding DCT transform. Local contrast enhancement is performed by nonlinear modification of denoised local DCT coefficients. To provide image processing in real time, a fast recursive algorithm for computing the sliding DCT is utilized. The algorithm is based on a recursive relationship between three subsequent local DCT spectra. Computer simulation results using a real image are provided and discussed.

This paper proposes a novel technique to extract range using a phase-only filter for a laser beam. The workspace is partitioned according to

M

meaningful preplanned range segments, each representing a relevant range segment in the scene. The phase-only filter codes the laser beam into

M

different diffraction patterns, corresponding to the predetermined range of each segment. Once the scene is illuminated by the coded beam, each plane in it would irradiate in a pattern corresponding to its range from the light source. Thus, range can be extracted at acquisition time. This technique has proven to be very efficient for qualitative real-time range extraction, and is mostly appropriate to handle mobile robot applications where a scene could be partitioned into a set of meaningful ranges, such as obstacle detection and docking. The hardware consists of a laser beam, a lens, a filter, and a camera, implying a simple and cost-effective technique.

A segmentation method based on a physics-based model of image formation is presented in this paper. This model predicts that, in image areas of uniform reflectance, colour channels keep coupled in the sense that they are not free to take any intensity value, but they depend on the values taken by other colour channels. This paper first enumerates and analyzes a set of properties in which this coupling materializes. Next, a segmentation strategy named C

3

S and based on looking for violations of the coupling properties is proposed. Segmentation results for synthetic and real images are presented at the end of the paper.

Watersheds are very powerful for image segmentation, and seeded watersheds have shown to be useful for object detection in images of cells in vitro. This paper shows that if cells are imaged over time, segmentation results from a previous time frame can be used as seeds for watershed segmentation of the current time frame. The seeds from the previous frame are combined with morphological seeds from the current frame, and over-segmentation is reduced by rule-based merging, propagating labels from one time-frame to the next. Thus, watershed segmentation is used for segmentation as well as tracking of cells over time. The described algorithm was tested on neural stem/progenitor cells imaged using time-lapse microscopy. Tracking results agreed to 71% to manual tracking results. The results were also compared to tracking based on solving the assignment problem using a modified version of the auction algorithm.

The task considered in this paper is performance evaluation of region segmentation algorithms in the ground truth (GT) based paradigm. Given a machine segmentation and a GT reference, performance measures are needed. We propose to consider the image segmentation problem as one of data clustering and, as a consequence, to use measures for comparing clusterings developed in statistics and machine learning. By doing so, we obtain a variety of performance measures which have not been used before in computer vision. In particular, some of these measures have the highly desired property of being a metric. Experimental results are reported on both synthetic and real data to validate the measures and compare them with others.

The paper describes a new image segmentation algorithm called

Combined Genetic segmentation

which is based on a genetic algorithm. Here, the segmentation is considered as a clustering of pixels and a similarity function based on spatial and intensity pixel features is used. The proposed methodology starts from the assumption that an image segmentation problem can be treated as a

Global Optimization Problem

. The results of the image segmentations algorithm has been compared with recent existing techniques. Several experiments, performed on real images, show good performances of our approach compared to other existing methods.

An effective method for the early detection of breast cancer is the mammographic screening. One of the most important signs of early breast cancer is the presence of microcalcifications. For the detection of microcalcification in a mammography image, we propose to conceive a multi-agent system based on a dual irregular pyramid.

An initial segmentation is obtained by an incremental approach; the result represents level zero of the pyramid. The edge information obtained by application of the Canny filter is taken into account to affine the segmentation. The edge-agents and region-agents cooper level by level of the pyramid by exploiting its various characteristics to provide the segmentation process convergence.

A new wavelet-domain HMTseg method is proposed, which fuses the segmentation results at coarse and fine scales with a new and feasible context model together with one preprocessing of raw segmentations at different scales. Compared to the original HMTseg method, the new method not only lays emphasis on the performance from coarse-scale segmentation, preserves the main outlines of the homogeneous regions in an image, and thus achieves good region consistency of segmentation, but also take into account the information from fine-scale segmentation, thus improves the accuracy of boundary localization of segmentation and enables the discrimination of small targets in an image, which is desirable for interpretation of remotely sensed images. Experiments on remotely sensed images, including aerial photos and SAR images, demonstrate that the proposed method can effectively take into consideration both the region consistency and the accuracy of boundary localization of segmentation performance, and give better segmentation results.

A novel method for segmentation of cardiac structures in temporal echocardiographic sequences based on the snake model is presented. The method is motivated by the observation that the structures of neighboring frames have consistent locations and shapes that aid in segmentation. To cooperate with the constraining information provided by the neighboring frames, we combine the template matching with the conventional snake model. Furthermore, in order to auto or semi-automatically segment the sequent images without manually drawing the initial contours in each image, generalized Hough transformation (GHT) is used to roughly estimate the initial contour by transforming the neighboring prior shape. As a result, the active contour can easily detect the desirable boundaries in ultrasound images.

We present a novel algorithm for binary image segmentation based on polygonal Markov fields. We recall and adapt the dynamic representation of these fields, and formulate image segmentation as a statistical estimation problem for a Gibbsian modification of an underlying polygonal Markov field. We discuss briefly the choice of Hamiltonian, and develop Monte Carlo techniques for finding the optimal partition of the image. The approach is illustrated by a range of examples.

Fingerprint image segmentation is one key step in Automatic Fingerprint Identification System (AFIS), and how to do it faster, more accurately and more effectively is important for AFIS. This paper introduces the Markov Chain Monte Carlo (MCMC) method and the Genetic Algorithm (GA) into fingerprint image segmentation and brings forward a fingerprint image segmentation method based on Markov Chain Monte Carlo and Genetic Algorithm (MCMC&GA). Firstly, it generates a random sequence of closed curves, which is regarded as the boundary between the fingerprint image region and the background image region, as Markov Chain, which uses boundary curve probability density function (BCPDF) as the index of convergence. Then, it is simulated by Monte Carlo method with BCPDF as a parameter, which is converged at the maximum. Lastly, Genetic Algorithm is introduced to accelerate the convergent speed. In conclusion, the closed curve with the maximum value of the BCPDF is the ideal boundary curve. The experimental results indicate that the method is robust to the low-quality finger images.

This paper proposes a method that aims to reduce a real scene to a set of regions that contain text fragments and keep small number of false positives. Text is modeled and characterized as a texture pattern, by employing the QMF wavelet decomposition as a texture feature extractor. Processing includes segmentation and spatial selection of regions and then content-based selection of fragments. Unlike many previous works, text fragments in different scales and resolutions laid against complex backgrounds are segmented without supervision. Tested in four image databases, the method is able to reduce visual noise to 4.69% and reaches 96.5% of coherency between the localized fragments and those generated by manual segmentation.

This paper introduces a new efficient way for computing affine invariant features from gray-scale images. The method is based on a novel image transform which produces infinitely many different invariants, and is applicable directly to isolated image patches without further segmentation. Among methods in this class only the affine invariant moments have as low complexity as our method, but as known they also possess many considerable weaknesses, including sensitivity to noise and occlusions. According to performed experiments it turns out that our novel method is more robust against these nonaffine distortions observed in image acquisition process, and even in practice its computation time is equivalent to that of the affine invariant moments. It is also observed that already a small subset of these new features is enough for successful classification.

The shapes of the objects in the images are important in the content-based image retrieval systems. In the contour-based shape description, Fourier descriptors have been proved to be effective and efficient methods. However, in addition to contour shape, Fourier description can be used to characterize also the color of the object. In this paper, we introduce new Color Fourier descriptors. In these descriptors, the boundary information is combined with the color of the object. The results obtained from the retrieval experiments show that by combining the color information with the boundary shape of the object, the retrieval accuracy can be clearly improved. This can be done without increasing the dimensionality of the descriptor.

This paper describes how facial shape can be modelled using a statistical model that captures variations in surface normal direction. We fit the model to intensity data using constraints on the surface normal direction provided by Lambert’s law. We demonstrate that this process yields improved facial shape recovery and can be used for the purposes of illumination insensitive face recognition.

This paper adopts face localization to eye extraction strategy for eye detection in complex scenes. First, an energy analysis is applied to enhance face localization performance by removing most noise-like regions rapidly. According to anthropometry, the face-of-interest (FOI) region is located using signatures derived from the proposed head contour detection (HCD) approach that searches the best combinations of facial sides and head contours. Second, via the de-edging preprocessing for facial sides, a wavelet subband inter-orientation projection method is devised to generate and select eye-like candidates. By utilizing the geometric discrimination information among the facial components, such as the eyes, nose, and mouth, the proposed eye verification rules verify the eye pair selected from the candidates. The experimental results demonstrate the significance performance improvement using the proposed method over others on three head-and-shoulder databases.

In several application projects we have discovered the need of computing the maximal inscribed convex set of a digital shape. Here we present an algorithm for computing a reasonable approximation of this set, that can be used in both 2D and 3D. The main idea is to iteratively identify the deepest concavity and then remove it by cutting off as little as possible of the shape. We show results using both synthetic and real examples.

This paper deals with the computation of homographies from two views in a multi-plane scene. In the general case, homographies can be determined using four matched points or lines belonging to planes. We propose an alternative method when a first homography has been obtained, and then three matches are sufficient to compute a second homography. This process is based on the geometric constraint introduced by the first homography. In this work, the extraction and matching of features, points or lines, is automatically performed using robust techniques. Experimental results with synthetic data and real images show the advantages of this approach. Besides, the performance using points or lines as image features is compared.

This paper describes a graph-spectral method for simplifying the structure of a graph. Our starting point is the lazy random walk on the graph, which is determined by the heat-kernel of the graph and can be computed from the spectrum of the graph Laplacian. We characterise the random walk using the commute time between nodes, and show how this quantity may be computed from the Laplacian spectrum using the discrete Green’s function. Our idea is to augment the graph with an auxiliary node which acts as a heat source. We use the pattern of commute times from this node to decompose the graph into a sequence of layers. These layers can be located using the Green’s function. We exploit this decomposition to develop a layer-by-layer graph-matching strategy. The matching method uses the commute time from the auxiliary node as a node-attribute.

Graphs have been successfully used in many disciplines of science and engineering. In the field of pattern recognition and image analysis, graph matching has proven to be a powerful tool. In this paper we generalize various matching tasks from graphs to the case of hypergraphs. We also discuss related algorithms for hypergraph matching.

In this paper, we explore whether the geometric properties of the point distribution obtained by embedding the nodes of a graph on a manifold can be used for the purposes of graph clustering. The embedding is performed using the heat-kernel of the graph, computed by exponentiating the Laplacian eigen-system. By equating the spectral heat kernel and its Gaussian form we are able to approximate the Euclidean distance between nodes on the manifold. The difference between the geodesic and Euclidean distances can be used to compute the sectional curvatures associated with the edges of the graph. To characterise the manifold on which the graph resides, we use the normalised histogram of sectional curvatures. By performing PCA on long-vectors representing the histogram bin-contents, we construct a pattern space for sets of graphs. We apply the technique to images from the COIL database, and demonstrate that it leads to well defined graph clusters.

The combination of multiple classifiers has been successful in improving classification accuracy in many pattern recognition problems. For graph matching, the fusion of classifiers is normally restricted to the decision level. In this paper we propose a novel fusion method for graph patterns. Our method detects common parts in graphs in an error-tolerant way using graph edit distance and constructs graphs representing the common parts only. In experiments, we demonstrate on two datasets that the method is able to improve the classification of graphs.

This paper describes work aimed at developing a practical scheme for face analysis using shape-from-shading. Existing methods have a tendency to recover surfaces in which convex features such as the nose are imploded. This is a result of the fact that subtle changes in the elements of the field of surface normals can cause significant changes in the corresponding integrated surface. To overcome this problem, in this paper we describe a local-shape based method for imposing convexity constraints. We show how to modifying the orientations in the surface gradient field using critical points on the surface and local shape indicators. The method is applied to both surface height recovery and face re-illumination, resulting in a clear improvement.

The implications of using commercially available non-collimated and distributed illuminates for the application of photometric stereo to dynamic surface inspection tasks are considered. A new form of lighting model, termed the virtual point light source model, is proposed for modelling real distributed illuminates in relative close proximity. The new technique has application for the two- and three-dimensional inspection of moving surfaces using an innovative technique known as dynamic photometric stereo. Such surface inspection tasks have previously been considered difficult or impossible to undertake using conventional imaging techniques. Experimental results are presented in the paper.

This paper investigates a kernel spectral approach to the problem of point pattern matching. Our first contribution is to show how kernel principal components analysis can be effectively used for solving the point correspondence matching problem when the point-sets are subject to structural errors, i.e. they are of different size. Our second contribution is to show how label consistency constraints can be incorporated into the construction of the Gram matrices for solving the articulated point pattern matching problem. We compare our algorithm with earlier point matching approaches and provide experiments on both synthetic data and real world data.

Shape-based recovery from image or video databases has become an important information retrieval problem. It is particularly challenging, owning to the difficulty to derive a similarity measurement that closely conforms to the common perception of humans. The goal of the current work is to achieve idea retrieval accuracy with reasonable speed and support for partial and occluded shapes. So, in this paper we introduce the elastic matching that is inspired by Duncan and Ayache combined with snake as a new shape retrieval technique. The elastic matching is to minimize of a quadratic fitting criterion, which consists of a curvature dependent bending energy term and a smoothness term. To reduce the computational complexity, the equation corresponding is only to the minimization of one-dimensional fitting criterion. As a result, the method proposed has the advantage of retrieve resemble objects with reasonable speed and less training samples.

In the fields of Augmented Reality (AR) and Virtual Reality (VR), inserting an object into a scene (real or virtual) requires proper matching of their lighting conditions. If not, the resulting image may look unnatural. In particular, it is important to describe the position and shape of specular reflection accurately if the object has specular reflection. In this paper, we propose an approach to relighting a moving object based on the separation of specular and diffuse reflection. To relight an object, two or more images taken under the condition that the position of the object is fixed but the lighting condition is different, we call

synchronized images

, are required. However, it is impossible to obtain such images in case of a moving object. Therefore, we propose a method that computationally obtains the synchronized images using the consecutive fields of a controlled video sequence containing a moving object. For example, if the virtual (

n

+ 1) –

th

field is interpolated from

n

–

th

field and (

n

+ 2) –

th

field using the motion compensation technique, both the virtual (

n

+ 1) –

th

field and the real (

n

+ 1) –

th

field have the condition that the position of the object is fixed. If the virtual and real image have different lighting condition, the method applied to static object is applicable to moving object as it is. After the specular and diffuse reflection are separted, the relit image is synthesized using the linear interpolation and morphing technique. The experimental results of applying the proposed method to real and synthetic images are given. We verify the effectiveness of the proposed method by comparing the resulting image with a ground-truth image.

This paper presents a new method of elastic articulated objects (human bodies) modeling based on a new conic curve. The model includes 3D object deformable curves which can represent the deformation of human occluding contours. The deformation of human occluding contour can be represented by adjusting only four deformation parameters for each limb. Then, the 3D deformation parameters are determined by corresponding 2D contours from a sequences of stereo images. The algorithm presented in this paper includes deformable conic curve parameters determination and the plane, 3D conic curve lying on, parameter determination.

This paper studies shape representation of general 3D objects. In particular, it proposes a conformal representation for genus-zero mesh objects, by using the discrete conformal mapping technique. It also proposes a new method to reconstruct the original shape from its conformal representation. In order to simplify and robustify the computation, we made several improvements to the above two procedures. The modifications include planar graph drawing initialization, Moebius factorization and spring-embedding-based reconstruction, etc. Though being mostly incremental, these modifications do provide significant improvements on previous methods. Possible applications include 3D geometry compression and object classification/recognition, etc.

Parameterized wavelet filters and wavelet packet subband structures are discussed to be used as key dependent wavelet transforms in order to enhance the security of wavelet based hashing schemes. Experiments show that key dependency and keyspace of the hashing scheme considered have been significantly improved. The attack resistance could only be slightly enhanced by using parametrized wavelet filters.

The 4×4 approximate discrete cosine transform (DCT) of H.264/AVC [1] makes it difficult to transcode the pre-coded video contents with the previous video coding standards to H.264/AVC in DCT domain. This is due to the difference between 8×8 DCT used previous standards and 4×4 DCT in H.264/AVC. In this paper, we propose an efficient algorithm that converts the quantized 8×8 DCT block of MPEG-2 into newly quantized four 4×4 DCT blocks of H.264/AVC to support DCT-domain transcoding. Experimental results show that the proposed scheme improves computational complexity by 5~11% and video quality by 0.1 ~ 0.5 dB compared with cascaded pixel-domain transcoding that exploits inverse quantization (IQ), inverse DCT (IDCT), DCT, and re-quantization (re-Q).

This paper presents an authentication algorithm based on robust watermarking techniques. In particular, the proposed method is based on a

self – embedding

scheme that is able not only to authenticate noisy images, but also to recover areas modified by a software pirate. The attack method investigated are semantic (altering the meaning of what the image is about) tampering, Gaussian white noise superposition, and JPEG2000 compression. The results are checked against the TAF function, which measure the distance between the inserted and the extracted watermark, and compared to similar algorithms in literature.

This paper presents a watermarking procedure for MPEG-2 videos based on the use of XML documents. The procedure enables the copyright owner to insert a distinct watermark identifying the buyer within the distributed videos. To increase the security level of the procedure, the watermark is repeatedly embedded into some selected I-frames in the DCT domain at different frequencies and by exploiting both block classification techniques and perceptual analysis. The embedded watermark is then extracted from a video according to the information contained in a protected XML document associated to the video.

To improve video quality and coding efficiency, H.264/AVC [1] has adopted several newer coding tools such as a 4×4 integer DCT and a method for calculating half pixels than the previous standards. However, these tools require additional work to transcode video content for pre-coded in the previous standards to H.264/AVC in the DCT domain. In this paper, we propose, as far as we know, the first half-pixel correction method for MPEG-2 to H.264 transcoding in the DCT domain. In the proposed method, an MPEG-2 block is added to the correction block obtained from the difference between half-pixel values of the two standards using a DCT reference frame. Experimental results show that the proposed method achieves better quality than the pixel based transcoding method.

Image processing applications are computing demanding and since a long time much attention has been paid to the use of parallel processing. Emerging distributed and Grid based architectures represent new and well suited platforms that promise the availability of the required computational power. In this direction image processing has to evolve to heterogeneous environments, and a crucial aspect is represented by the interoperability and reuse of available and high performance code. This paper describes our experience in the development of PIMA(GE)

2

, Parallel IMAGE processing GEnoa server, obtained wrapping a library using the CORBA framework. Our aim is to obtain a high level of flexibility and dynamicity in the server architecture with a possible limited overhead. The design of a hierarchy of image processing operation objects and the development of the server interface are discussed.

This paper shows how Markovian segmentation algorithms used to solve well known computer vision problems such as

motion estimation

,

motion detection

and

stereovision

can be significantly accelerated when implemented on programmable graphics hardware. More precisely, this contribution exposes how the parallel abilities of a standard Graphics Processing Unit (usually devoted to image synthesis) can be used to infer the labels of a label field. The computer vision problems addressed in this paper are solved in the maximum a posteriori (MAP) sense with an optimization algorithm such as ICM or simulated annealing. To do so, the

fragment processor

is used to update in parallel every labels of the segmentation map while rendering passes and graphics textures are used to simulate optimization iterations. Results show that impressive acceleration factors can be reached, especially when the size of the scene, the number of labels or the number of iterations is large. Hardware results have been obtained with programs running on a mid-end affordable graphics card.

At the low compression ratio, the EBCOT engine of the JPEG 2000 encoder does not have to process all input data to achieve an optimal codestream in the sense of the rate-distortion criteria. This property is exploited in the architecture presented in this paper to allow higher throughputs of the JPEG 2000 encoder. An impact of the code block size and the internal FIFO size on the resultant speed is considered. The architecture is described in VHDL and synthesized for commercial FPGA technology. Simulation results show that at low compression ratios and for FPGA Stratix II devices, the single engine can support HDTV standards.

In this paper, we improve the performance of a generic visual categorizer based on the ”bag of keypatches” approach using geometric information. More precisely, we consider a large number of simple geometrical relationships between interest points based on the scale, orientation or closeness. Each relationship leads to a weak classifier. The boosting approach is used to select from this multitude of classifiers (several millions in our case) and to combine them effectively with the original classifier. Results are shown on a new challenging 10 class dataset.

The aim of the present work is building an evaluation method for the similarity between colour hues. The method is defined by studying the attribution process, by human subjects, of colour hue couple to similarity classes (from ‘very similar’ to ‘little similar’). From the study of these categorical judgements it is derived that the relation between the hue and the colour similarity is ‘not-isometric’ and greatly depends on the colour category. This result allows to extract representative functions for the three colour of the subtractive system: Red, Yellow, Blue. Besides we used a new method for segmenting the colour, based on the similarity with the main colours. Our method defines a quaternary tree structure, named ‘Similarity Quad-Tree’; it is capable of extracting, from the whole image, the belonging degree to the Red, Yellow and Blue colours and their similarity with the reference colour. The check on the method applicability has given good results both: in the user satisfaction and in the computation. The approach may be viewed as a simple and fast indexing method.

An efficient approach to shape matching in digital images is presented. The method, called Weighted Edge Potential Function, is a significant improvement of the EPF similarity measure, which models the image edges as charged elements in order to generate a field of attraction over similarly shaped objects. Experimental results and comparisons demonstrate that WEPF enhances the properties of EPF and outperforms traditional similarity metrics in shape matching applications, in particular in the presence of noise and clutter.

In this paper, we present an automatic system that is able to forecast the appearance of a soccer highlight, and annotate it, based on MPEG features; processing is performed in strict real time. A probabilistic framework based on Bayes networks is used to detect the most significant soccer highlights. Predictions are validated by different Bayes networks, to check the outcome of forecasts.

A lightweight encryption scheme for visual data based on wavelet filter parametrization is discussed. Being a special variant of header encryption, the technique has an extremely low computational demand. Security assessement of low-dimensional parametrizations schemes show severe weaknesses. We show that using high-dimensional parametrizations the scheme may be employed in applications requiring a medium security level.

In this paper we describe general algorithms that are devised for MPEG-7 automatic annotation of Formula 1 videos, and in particular for camera-car shots detection. We employed a shot detection algorithm suitable for cuts and linear transitions detection, which is able to precisely detect both the transition’s center and length. Statistical features based on MPEG motion compensation vectors are then employed to provide motion characterization, using a subset of the motion types defined in MPEG-7, and shot type classification. Results on shot detection and classification are provided.

This paper focuses on soccer ball tracking which is known to be more difficult than that of players due to its small size in an image and abrupt changes in its motion. Suggested is an effective soccer ball tracking algorithm which estimates ball position by exploiting the background image and player tracking results. In other words, the trajectory of ball is derived as image blobs by eliminating player blobs and the background parts from an image sequence. This algorithm performed well on a pretty long TV broadcast sequence in which the ball is frequently occluded by players.

The Georgia Department of Transportation (GDOT) collects and maintains an inventory of all public roads within the state. The inventory covers more than 118,000 centerline miles (188,800 km) of roads in 159 counties and over 512 municipalities. The transportation road inventory includes more than 52 items, including roadway geometry, surface type, shoulder type, speed limit signs, etc. Traditional roadway geometric properties, including number of lanes, travel lane, and shoulder widths, are measured in the field. Roadway geometric property measurement is one of the most important and, yet, the most time-consuming and riskiest component of the roadway data inventory. For the past two years, GDOT has sponsored Georgia Tech to develop a GPS/GIS-based road inventory system that re-engineers the existing paper-pencil operations. Georgia Tech has extended the research to develop video image pattern recognition algorithms and a prototype application aimed at automating the roadway geometry measurement to enhance the roadway inventory operations. A highly reliable and effective image extraction algorithm using local thresholding, predictive edge extraction, and geometric optics was developed and is presented in this paper. Preliminary results show it can effectively extract roadway features. A large-scale, experimental study on accuracy and the productivity improvement is under way.

Segmentation of news videos into stories is among key issues for achieving efficient treatment of news-based digital libraries. Indeed, anchor shot detection is a fundamental step for segmenting news into stories.

In this paper we present an improved algorithm for anchor shot detection. It is based on a graph theoretical clustering method and exploits the idea of the

cluster lifetime

for improving the performance. Moreover, a method for automatically fixing all the thresholds required by the original version of the algorithm is also proposed, so making it fully unsupervised.

The proposed algorithm has been tested on a database significantly wider than those typically used in the field, demonstrating its advantages with respect to the original version.

In this note it is discussed how face detection and tracking in video can be achieved relying on a

detection-tracking loop

. Such integrated approach is appealing with respect either to robustness and computational efficiency.

Many algorithms have been proposed in literature for digital film restoration; unfortunately, none of them ensures a perfect restoration whichever is the image sequence to be restored. Here, we propose an approach to digital scratch restoration based on image fusion techniques for combining relatively well settled distinct techniques. Qualitative results are deeply investigated for several real image sequences.

Automatic recognition of time and date stamps in CCTV video enables the inclusion of time-based queries in video indexing applications. Such ability needs to deal with problems of low character resolution, non-uniform background, multiplexed video format, and random access to the video file. In this paper, we address these problems and propose a technique that solves the difficult task of character segmentation, by means of a recognition-based process. Our method consists of three main steps: pattern matching, character location and syntactic analysis. The experiments prove the reliability and efficiency of the proposed method, obtaining an overall recognition rate over 80%.

An innovative algorithm for automatic generation of Huffman coding tables for semantic classes of digital images is presented. Collecting statistics over a large dataset of corresponding images, we generated Huffman tables for three images classes: landscape, portrait and document. Comparisons between the new tables and the JPEG standard coding tables, using also different quality settings, have shown the effectiveness of the proposed strategy in terms of final bit size (e.g. compression ratio).

This paper presents an efficient implementation of moving least square(MLS) approximation for 3D surface reconstruction. The smoothness of the MLS is mainly determined by the weight function where its support greatly affects the accuracy as well as the computational time in the mixed dense and scattered data. In a point-set, possibly acquired from a 3D scanning device, it is important to determine the support of the weight function adaptively depending on the distribution and shape of the given scatter data. Particulary in case of face data including the very smooth parts, detail parts and some missing parts of hair due to low reflectance, preserving some details while filling the missing parts smoothly is needed. Therefore we present a fast algorithm to estimate the support parameter adaptively by a raster scan method from the quantized integer array of the given data. Some experimental results show that it guarantees the high accuracy and works to fill the missing parts very well.

A new genetic programming based approach to classification problems is proposed. Differently from other approaches, the number of prototypes in the classifier is not a priori fixed, but automatically found by the system. In fact, in many problems a single class may contain a variable number of subclasses. Hence, a single prototype, may be inadequate to represent all the members of the class. The devised approach has been tested on several problems and the results compared with those obtained by a different genetic programming based approach recently proposed in the literature.

We describe an automatic procedure for building risk maps of unexploded ordnances (UXO) based on historic air photographs. The system is based on a cost-sensitive version of AdaBoost regularized by hard point shaving techniques, and integrated by spatial smoothing. The result is a map of the spatial density of craters, an indicator of UXO risk.

This paper presents a system for classifying facial expressions based on a data fusion process relying on the Belief Theory (BeT). Four expressions are considered:

joy, surprise, disgust

as well as

neutral

. The proposed system is able to take into account intrinsic doubt about emotion in the recognition process and to handle the fact that each person has his/her own maximal intensity of displaying a particular facial expression. To demonstrate the suitability of our approach for facial expression classification, we compare it with two other standard approaches: the Bayesian Theory (BaT) and the Hidden Markov Models (HMM). The three classification systems use characteristic distances measuring the deformations of facial skeletons. These skeletons result from a contour segmentation of facial permanent features (mouth, eyes and eyebrows). The performances of the classification systems are tested on the Hammal-Caplier database [1] and it is shown that the BeT classifier outperforms both the BaT and HMM classifiers for the considered application.

In this paper, we propose a novel method which involves neural adaptive techniques for identifying salient features and for classifying high dimensionality data. In particular a network pruning algorithm acting on MultiLayer Perceptron topology is the foundation of the feature selection strategy. Feature selection is implemented within the back-propagation learning process and based on a measure of saliency derived from bell functions positioned between input and hidden layers and adaptively varied in shape and position during learning. Performances were evaluated experimentally within a Remote Sensing study, aimed to classify hyperspectral data. A comparison analysis was conducted with Support Vector Machine and conventional statistical and neural techniques. As seen in the experimental context, the adaptive neural classifier showed a competitive behavior with respect to the other classifiers considered; it performed a selection of the most relevant features and showed a robust behavior operating under minimal training and noisy situations.

We investigate accuracy, neural network complexity and sample size problem in multilayer perceptron (MLP) based (neuro-linear) feature extraction. For feature extraction we use weighted sums calculated in hidden units of the MLP based classifier. Extracted features are utilized for data visualisation in 2D and 3D spaces and interactive formation of the pattern classes. We show analytically how complexity of feature extraction algorithm depends on the number of hidden units. Sample size – complexity relations investigated in this paper showed that reliability of the neuro-linear feature extraction could become extremely low if number of new features is too high. Visual interactive inspection of data projection may help an investigator to look differently at the forecasting problem of the financial time series.

Combining classifiers has proved to be an effective solution to several classification problems in pattern recognition. In this paper we use classifier combination methods for the classification of natural images. In the image classification, it is often beneficial to consider each feature type separately, and combine the classification results in the final classifier. We present a classifier combination strategy that is based on classification result vector, CRV. It can be applied both in supervised and unsupervised manner. In this paper we apply our classifier combination method to the classification of rock images that are non-homogenous in terms of their color and texture properties.

A well established technique to improve the classification performances is to combine more classifiers. In the binary case, an effective instrument to analyze the dichotomizers under different class and cost distributions providing a description of their performances at different operating points is the Receiver Operating Characteristic (ROC) curve. To generate a ROC curve, the outputs of the dichotomizers have to be processed. An alternative way that makes this analysis more tractable with mathematical tools is to use a parametric model and, in particular, the binormal model that gives a good approximation to many empirical ROC curves. Starting from this model, we propose a method to estimate the ROC curve of the linear combination of two dichotomizers given the ROC curves of the single classifiers. A possible application of this approach has been successfully tested on real data set.

Techniques for prediction in spatial maps can be based on associative neural network models. Unfortunately, the performance of standard associative memories depends on the number of training patterns stored in the memory; moreover it is very sensitive to mutual correlations of the stored patterns. In order to overcome limitations imposed by processing of a large number of mutually correlated spatial patterns, we have designed the Hierarchical Associative Memory model which consists of arbitrary number of associative memories hierarchically grouped into several layers. In order to further improve its recall abilities, we have proposed new modification of our model. In this paper, we also present experimental results focused on recall ability of designed model and their analysis by means of mathematical statistics.

The problem of feature points matching between pair of views of the scene is one of the key problems in computer vision, because of the number of applications. In this paper we discuss an alternative version of an SVD matching algorithm earlier proposed in the literature. In the version proposed the original algorithm has been modified for coping with large baselines. The claim of improved performances for larger baselines is supported by experimental evidence.

The paper refers an industrial solution to 3D measurements to be used in security application. The basic assumption of the proposed solution is the visibility of a 3D reference system in the imaged scene. The corresponding vanishing points are used as stereo point pairs at infinity, to compute the rotation between the two stereo views. To achieve a metric reconstruction, the proposed approach relies on the presence in the scene of an artificial landmark having an orthogonal triangular shape with sides of one meter length.. Reflecting cylinders are positioned at the corners and along the sides at equally spaced distances. They represent the initial stereo correspondence for system calibration. This approach has been implemented in a prototype product for the planar reconstruction in case of car accidents using a digital camera. The software application provides a simple and user-friendly interface to the process. The measure errors are always within 2 % (with a 4Mp camera).

SSSD-based linear multibaseline stereo is an efficient implementation of multi-camera stereo vision system. This efficiency, however, vitally relies on the ideal configuration of all cameras. For dealing with non-ideal configurations, conventional stereo rectification algorithms can be used, but the performances are often still not satisfactory. This paper proposes a new algorithm to process non-ideally configured multibaseline stereo system, which not only avoids the rectification procedure but also remains the efficiency of SSSD at the same time. This is fulfilled by using the idea of tensor transfer used in image-based-rendering area. In particular, the multibaseline stereo is reformulated as a novel-view-synthesis problem. We propose a new concept of tensor-transfer to generate novel views as well as compute the depth map.

This paper presents an optimisation technique to select automatically a set of control parameters for a Markov Random Field applied to stereo matching. The method is based on the Reactive Tabu Search strategy, and requires to define a suitable fitness function that measures the performance of the MRF stereo algorithm with a given parameters set. This approach have been made possible by the recent availability of ground-truth disparity maps. Experiments with synthetic and real images illustrate the approach.

A new dynamic photometric stereo technique is presented, which facilitates the analysis of rapidly moving surfaces containing mixed two- and three-dimensional concomitant features. A new approach termed narrow infra-red photometric stereo (NIRPS) is described as an evolution of the existing photometric stereo (PS) technique. The method has application for the inspection of web materials and other moving surfaces considered difficult to analyse using conventional imaging techniques. Experimental results are presented in the paper.

Solids of revolution (vases, bottles, bells, ...), shortly SORs, are very common objects in man-made environments. We present a complete framework for 3D database population from a single image of a scene including a SOR. The system supports human intervention with automated procedures to obtain good estimates of the interest entities needed for 3D reconstruction. The system exploits the special geometry of the SOR to localize it within the image and to recover its shape. Applications for this system range from the preservation and classification of ancient vases to advanced graphics and multimedia.

In this paper we present a 3D human face reconstruction framework based on stereo sensor coupled with a structured lighting source. Starting from two calibrated images, the active part (video projector) which project controlled lights, allows the operator to locate two sets of structured features with sub-pixel accuracy in both left and right images. Then, exploiting epipolar geometry improves the matching process by reducing its complexity from a bidirectional to a unidirectional search problem. Finally, we perform an adapted dynamic programming algorithm to obtain corresponding features in each conjugated scanline separately. Final three dimensional face models are achieved by a pipeline of four steps: (a) stereo triangulation, (b) data interpolation based on cubic spline models, (c) Delaunay triangulation-based meshing, and (d) texture mapping process.

We are researching for real-time hand shape estimation, which we are going to apply to user interface and interactive applications. We have employed a computer vision approach, since unwired sensing provides restriction-free observation, or a natural way of sensing. The problem is that since a human hand has many joints, it has geometrically high degrees of freedom, which makes hand shape estimation difficult. For example, we have to deal with a self-occlusion problem and a large amount of computation. At the same time, a human hand has several physical constraints, i.e., each joint has a movable range and interdependence, which can potentially reduce the search space of hand shape estimation. This paper proposes a novel method to estimate 3D hand shapes in real-time by using shape features acquired from camera images and physical hand constraints heuristically introduced. We have made preliminary experiments using multiple cameras under uncomplicated background. We show experimental results in order to verify the effectiveness of our proposed method.

Along with images and videos, 3D models have raised a certain interest for a number of reasons, including advancements in 3D hardware and software technologies, their ever decreasing prices and increasing availability, affordable 3D authoring tools, and the establishment of open standards for 3D data interchange. The resulting proliferation of 3D models demands for tools supporting their effective and efficient management, including archival and retrieval.

In order to support effective retrieval by content of 3D objects and enable retrieval by object parts, information about local object structure should be combined with spatial information on object surface. In this paper, as a solution to this requirement, we present a method relying on curvature correlograms to perform description and retrieval by content of 3D objects.

Experimental results are presented both to show results of sample queries by content and to compare—in terms of precision/recall figures—the proposed solution to alternative techniques.

We present a new scheme for 3D surface reconstruction of a moving object in the presence of specular reflection. We basically search for the depth at each point on the surface of the object while exploiting the recently proposed geotensity constraint [7] that accurately governs the relationship between four or more images of a moving object in spite of the illumination variance due to object motion. The thrust of this paper is then to extend the availability of the geotensity constraint to the case that specularities are also present. The key idea is to utilise the fact that highlights shift on the surface due to object motion. I.e., we employ five or more images as inputs, and interchangeably utilise a certain intensity subset consisting of four projected intensities which is the least influenced by the specular component. We illustrate the relevancy of our simple algorithm also through experiments.

This paper makes two contributions. First, we present an experimental analysis of three different ways of constructing three-dimensional statistical models of faces using Cartesian coordinates, namely, height, surface gradient and one based on Fourier domain basis functions. Second, we test the ability of each of the models for dealing with information provided by shape-from-shading by introducing a simple non-exhaustive parameter adjustment procedure subject to integrability and irradiance constraints. Experiments show that the surface gradient based representation is more robust to noise than the alternative Cartesian representations.

Mammography is a not invasive diagnostic technique widely used for early cancer detection in women breast. A particularly significant clue of such disease is the presence of clusters of microcalcifications. The automatic detection of such clusters is a very difficult task because of the small size of the microcalcifications and of the poor quality of the digital mammograms. In literature, all the proposed method for the automatic detection focus on the single microcalcification. In this paper, an approach that moves the final decision on the regions identified by the segmentation in the phase of clustering is proposed. To this aim, the output of a classifier on the single microcalcifications is used as input data in different clustering algorithms which produce the final decision. The approach has been successfully tested on a standard database of 40 mammographic images, publicly available.

Holter signals are ambulatory long-term electrocardiographic (ECG) registers used to detect heart diseases which are difficult to find in normal ECG. These signals normally include several channels and its duration is up to 48 hours. The principal problem for the cardiologists consists of the manual inspection of the whole Holter ECG to find all those beats whose morphology differ from the normal cardiac rhythm. The later analysis of these abnormal beats yields a diagnostic from the pacient’s heart condition. In this paper we compare the performance among several clustering methods applied over the beats processed by Principal Component Analysis (PCA). Moreover, an outlier removing stage is added, and a cluster estimation method is included. Quality measurements, based on ECG labels from MIT-BIH database, are developed too. At the end, some results-accuracy values among several clustering algorithms is presented.

Two-dimensional (2–

D

) autoregressive (AR) models have been used as one of the methods to characterise the textures of tumours in mammograms. Previously, the 2–

D

AR model coefficients were estimated for the block containing the tumour and the blocks in its 3 × 3 neighbourhood. In this paper, the possibility of having the estimated set of AR model coefficients of the block containing the tumour as a unique set of AR model coefficients for the entire mammogram is looked into. Based on the information given from the MiniMammography database, the possible number of blocks of the same size of the block containing the tumour is obtained from the entire mammogram and for each block a set of AR model coefficients is estimated using a method that combines both the Yule-Walker system of equations and the Yule-Walker system of equations in the third-order statistical domain. These sets of AR model coefficients are then compared. The simulation results show that 98.6% of the time we can not find another set of AR model coefficients representing the blocks of pixels in the possible neighbourhood of the entire mammogram for the data (95 mammograms with 5 of them having two tumours) available in the MiniMammography database.

In this paper we present a textural feature analysis applied to a medical image segmentation problem where other methods fail, i.e. the localization of thrombotic tissue in the aorta. This problem is extremely relevant because many clinical applications are being developed for the computer assisted, image driven planning of vascular intervention, but standard segmentation techniques based on edges or gray level thresholding are not able to differentiate thrombus from surrounding tissues like vena, pancreas having similar HU average and noisy patterns [3,4]. Our work consisted in a deep analysis of the texture segmentation approaches used for CT scans, and on experimental tests performed to find out textural features that better discriminate between thrombus and other tissues. Found that some Run Length codes perform well both in literature and experiments, we tried to understand the reason of their success suggesting a revision of this approach with feature selection and the use of specifically thresholded Run Lengths that improves the discriminative power of measures reducing the computational cost.

In this paper a linear transformation, the spiral method, is introduced; this transformation maps an image represented in a 3-D space into a signal in a 2-D space. Some features of the spiral are presented: for instance the topologic information of the objects in the image, their contours, areas and the shape of the objects themselves. Two different case-study are presented: the use of spiral method in order to evaluate the number, the size, the shape and the location of the microcalcifications by the use of signals related to the mammograms; entropy is proposed as a measure of the degree of the parenchyma disorder of the mammograms and its use for a system CAD.

A very important artifact corrupting Magnetic Resonance (MR) Images is the RF inhomogeneity, also called Bias artifact. The visual effect produced by this kind of artifact is an illumination variation which afflicts this kind of medical images. In literature a lot of works oriented to the suppression of this artifact can be found. The approaches based on homomorphic filtering offer an easy way to perform bias correction but none of them can automatically determine the cut-off frequency. In this work we present a measure based on information theory in order to find the frequency mentioned above and this technique is applied to MR images of the knee which are hardly bias corrupted.

In this paper, we propose a novel technique of registration using hybrid approach for MR-PET brain image fusion. Hybrid approach uses merits of surface- and voxel-based registration. Thus, our method measures similarities using voxel intensities in MR images corresponding to the feature points of the brain in PET images. Proposed method selects the brain threshold using histogram accumulation ratio in PET images. And then, we automatically segment the brain using the inverse region growing with pre-calculated threshold and extract the feature points of the brain using sharpening filter in PET images. In order to find the optimal location for registration, we evaluate the Hybrid-based Cross-Correlation using the voxel intensities in MR images corresponding to the feature points in PET images. In our experiments, we evaluate our method using software phantom and clinical datasets in the aspect of visual inspection, accuracy, robustness, and computation time. Experimental results show that our method is dramatically faster than the voxel-based registration and more accurate than the surface-based registration. In particular, our method can robustly align two datasets with large geometrical displacement and noise at optimal location.

Information systems are essential tools supporting the management of hospital organizations. The demand for availability and integration of data in this field is more and more increasing, basically for absolving two key issues: collecting and merging all the elementary data available for a patient during the hospitalization process, so that physicians and other operators get all the necessary information they need during the process; planning the development of the diagnostic activities/therapeutics to optimize the process. In this perspective, we present a system that integrates a booking subsystem for hospital specialized treatments booking (CUP), a subsystem for the management of the hospitalizations (including First Aid Departments), a subsystem for filing and reporting clinical images, a subsystem for the analysis of radiological images in a unique management environment. Therefore we describe a complete system for the management of an archive of digital dossiers for the patients of a hospital, where diagnostic imaging is a central issue.

The aim of the ESCALE project is to supply the medical staff with objective and accurate 2D and 3D measurements for wound healing assessment from color images acquired in a free manner with a low cost digital camera. The problem addressed in this paper is the volume estimation from uncalibrated views. We present two experimentations. A Monte Carlo simulation on synthetic perturbated data leads to an average error of 3% on reconstructed points. Then, triangulation based volume estimation obtained from two uncalibrated real images gives us hope that an accuracy less than 5% is achievable. So this technique is suited to accurate wound 3D measurements. Combined with true color image processing for colorimetric tissue assessment, a such low cost system will be appropriate tool for diagnosis assistance and therapy monitoring in clinical environment.

This paper presents the scatter search particle filter (SSPF) algorithm and its application to real-time hands and face tracking. SSPF combines sequential Monte Carlo (particle filter) and combinatorial optimization (scatter search) methods. Hands and face are characterized using a skin-color model based on explicit RGB region definition. The hybrid SSPF approach enhances the performance of classical particle filter, reducing the required evaluations of the weighting function and increasing the quality of the estimated solution. The system operates on 320x240 live video in real-time.

Most of the already proposed skin detection approaches are based on the same pixel-wise paradigm, in which each image pixel is individually analyzed. We think that this paradigm should be extended; context information should be incorporated in the skin detection process. Following this idea, in this article is proposed a robust and fast skin detection approach that uses spatial and temporal context. Spatial context implies that the decision about the class (skin or non-skin) of a given pixel considers information about the pixel’s neighbors. Temporal context implies that skin detection is carried out considering not only pixel values from the current frame, but also taking into account past frames and general background reference information.

This paper presents a new iris database that contains images with noise. This is in contrast with the existing databases, that are noise free. UBIRIS is a tool for the development of robust iris recognition algorithms for biometric proposes.

We present a detailed description of the many characteristics of UBIRIS and a comparison of several image segmentation approaches used in the current iris segmentation methods where it is evident their small tolerance to noisy images.

Spatial dimension reduction called Two Dimensional PCA method has recently been presented. The application of this variation of traditional PCA considers images as 2D matrices instead of 1D vectors as other dimension reduction methods have been using. The application of these advances to verification techniques, using SVM as classification algorithm, is here shown. The simulation has been performed over a complete facial images database called FRAV2D that contains different sets of images to measure the improvements on several difficulties such as rotations, illumination problems, gestures or occlusion.

The new method endowed with a classification strategy of SVMs, seriously improves the results achieved by the traditional classification of PCA & SVM.

Facial image analysis is very useful in many applications such as video compression, talking heads, or biometrics. During the last few years, many algorithms have been proposed in particular for face recognition using classical 2-D images. Face is fairly easy to use and well accepted by people but generally not robust enough to be used in most practical security applications because too sensitive to variations in pose and illumination. One possibility to overcome this limitation is to work in 3-D instead of 2-D. But 3-D is costly and more difficult to manipulate and then ineffective to authenticate people in most contexts. Hence, to solve this problem, we propose a novel face recognition approach that is based on an asymmetric protocol: enrolment in 3-D but identification performed from 2-D images. So that, the goal is to make more robust face recognition while keeping the system practical. To make this 3-D/2-D approach possible, we introduce geometric invariants used in computer vision within the context of face recognition. We report preliminary experiments to evaluate robustness of invariants according to pose variations and to the accuracy of detection of facial feature points. Preliminary results obtained in terms of identification rate are encouraging.

In this paper, an algorithm for extracting three-dimension shape of human face from two 2D images using grid light is presented. The grid pattern is illuminated by incandescence light instead of laser in order to protect human eyes or skin and reduce cost

.

An uncoded grid pattern is projected on human face to solve the problem of correspondence between a pair of stereo images. Two images acquired at same time are smoothed to diminish noise at first. Then grid stripes from these images are extracted and thinned by a marked watershed algorithm. A new method based on graph connectivity to locate and label grid intersections from these images is also presented. According to labeling principles, a set of matched points is build. The set of matched points are further used to calculate three-dimension-depth information of human face. Experiment results show the feasibility of the proposed method.

In this paper we present a general technique for face and facial feature localization in 2D color images with arbitrary background. In a previous work we studied an eye localization module, while here we focus on mouth localization. Given in input an image that depicts a sole person, first we exploit the color information to limit the search area to candidate mouth regions, then we determine the exact mouth position by means of a SVM trained for the purpose. This component-based approach achieves the localization of both the faces and the corresponding facial features, being robust to partial occlusions, pose, scale and illumination variations. We report the results of the separate modules of the single feature classifiers and their combination on images of several public databases.

We present in this paper a method for the localization of the eyes in a facial image. This method works on color images, applying the so called Chinese Transformation (CT) on edge pixels to detect local symmetry. The CT is combined with a skin color model based on a modified Gaussian Mixture Model (GMM). The CT and the modified GMM give us a small rectangular area containing one eye with a very high probability. This rectangle is then processed to find the precise position of the eye, using four sources of information: a darkness measure, a circle finder, a “not skin” finder and a position information. Experimental results on a large database are presented on nearly 1000 faces from the ECU database.

Fingerprints are one of the most used biometrics for automatic personal authentication. Unfortunately, it is often difficult to design fingerprint matchers exhibiting the performances required in real applications. To meet the application requirements, fusion techniques based on multiple matching algorithms, multiple fingerprints, and multiple impressions of the same fingerprint, have been investigated. However, no previous work has investigated selection strategies for biometrics. In this paper, a score selection strategy for fingerprint multi-modal authentication is proposed. For each authentication task, only one score is dynamically selected so that the genuine and the impostor users’ scores distributions are mainly separated. Score selection is performed by first estimating the likelihood that the input pattern is an impostor or a genuine user. Then, the min score is selected in case of an impostor, while the max score is selected in case of a genuine user. Reported results show that the proposed selection strategy can provide better performances than those of commonly used fusion rules.

In order for a subspace projection based method to be robust to local distortion and partial occlusion, the basis images generated by the method should exhibit a part-based local representation. We propose an effective part-based local representation method using ICA architecture I basis images that is robust to local distortion and partial occlusion. The proposed representation only employs locally salient information from important facial parts in order to maximize the benefit of applying the idea of “recognition by parts.” We have contrasted our representation with other part-based representations such as LNMF (Localized Non-negative Matrix Factorization) and LFA (Local Feature Analysis). Experimental results show that our representation performs better than PCA, ICA architecture I, ICA architecture II, LFA, and LNMF methods, especially in the cases of partial occlusions and local distortions.

Combining different algorithms submitted to the Third International Fingerprint Verification Competition (FVC2004) is studied. For this work, the matching results of more than 40 fingerprint systems from both academy and industry are available on standard benchmark data. In particular, we concentrate on score-level fusion of the different algorithms, studying the correlation between them by using feature-subset-selection techniques. Based on the general algorithmic descriptions provided by the participants, some interesting experimental findings are obtained.

We present a new method dedicated to the localization of faces in color images. It combines a connexionist model (auto-associative network), an ellipse model based on Generalized Hough Transform, a skin color model and an eyes detector that results in two features. A linear combination of the 3 first models is performed to eliminate most of non face regions. A connexionist combination of the four detectors response is performed on the remaining candidates. Given an input image, we compute a kind of probability map on it with a sliding window. The face position is then determined as the location of the absolute maximum over this map. Improvement of baseline detectors localization rates is clearly shown and results are very encouraging.

3D face recognition has gained growing attention in the last years, mainly because both the limitations of 2D images and the advances in 3D imaging sensors. This paper proposes a novel approach to perform 3D face matching by using a new metric, called the Surface Interpenetration Measure (SIM). The experimental results include a comparison with a state-of-art work presented in the literature and show that the SIM is very discriminatory as confronted with other metrics. The experiments were performed using two different databases and the obtained results were quite similar, showing the robustness of our approach.

This paper describes a method to detect and identify the typical italian road sign passo-carrabile. The system first determines the region of interest within the image, using color segmentation, then the signal of no waiting is identified using shape and color information, and finally the text label passo-carrabile is recognised with a state diagram and a set of specific tests on the shape of the words. The obtained results show the feasibility of the system.

Image warping caused by scanning, photocopying or photographing a document is a common problem in the .eld of document processing and understanding. Distortion within the text documents impairs OCRability and thus strongly decreases the usability of the results. This is one of the major obstacles for automating the process of digitizing printed documents.

In this paper we present a novel algorithm which is able to correct document image warping based on the detection of distorted text lines. The proposed solution is used in a recent project of digitizing old, poor quality manuscripts. The algorithm is compared to other published approaches. Experiments with various document samples and the resulting improvements of the text recognition rate achieved by a commercial OCR engine are also presented.

Image compositing is defined as the assembling of two or more overlapping images into a single image. Recently, a morphological image compositing algorithm was proposed that automatically positions seam lines along salient image structures. This algorithm requires that all images used in the compositing procedure be held in computer memory. Hence, when composing large images such as satellite imagery acquired over a large region, the maximal capacity of random access memory of 32 bit computers is rapidly exceeded. In this paper, we present a parallel algorithm producing the same results whilst requiring only one input image at a time to be held in memory. The algorithm is illustrated for the automatic production of pan-European mosaics of Landsat images.

In this article we proposed an improved SIFT-based object recognition methodology for robot applications. This methodology is employed for implementing a robot-head detection system, which is the main component of a robot gaze direction determination system. Gaze direction determination of robots is an important ability to be developed. It can be used for enhancing cooperative and competitive skills in situations where the robots interacting abilities are important, as for example, robot soccer. Experimental results of the implemented robot-head detection system are presented.

Robot navigation using only abstract, topological information on the environment is strongly related to the possibility for a robot to unambiguously match information coming from its sensors with the basic elements of the environment. In this paper we present an approach to this challenging problem based on the direct recognition of the topological structure of the environment.

In this research the problem of the automatic detection and classification of rectangular road sign has been faced. The first step concerns the robust identification of the rectangular sign, through the search of gray level discontinuity on the image and Hough transform. Due to variety of rectangular road signs, we first recognize the guide sign and then we consider advertising the other rectangular signs. The classification is based on analysis of surface color and arrows direction of the sign. We have faced different problems, primarily: shape alterations of the sign owed to the perspective, shades, different light conditions, occlusion. The obtained results show the feasibility of the system.

Recently, appearance-based approaches have attracted the interests of computer vision researchers. Based on this idea, an appearance-based navigation method using the View-Sequenced Route-Representation model is proposed. A couple of parallel cameras is used to take low-resolution fontal images along the route to follow. Then, zero mean cross correlation is used as image comparison criterion to perform auto-location and control of the robot using only visual information. Besides, a sensibility analysis of the navigation parameters has been carried out to try to optimize the accuracy and the speed in route following. An exhaustive number of experiments using a 4-wheel drive robot with synchronous drive kinematics have been carried out.

The paper deals with the use of Support Vector Machines (SVMs) and performance comparisons with Artificial Neural Networks (ANNs) in software-based Instrument Fault Accommodation schemes. As an example, a real case study on an automotive systems is presented. The ANNs and SVMs regression capability are employed to accommodate faults that could occur on main sensors involved in the operating engine. The obtained results prove the good behaviour of both tools and similar performances have been achieved in terms of accuracy.

Edit Distance has been widely studied and successfully applied in a large variety of application domains and many techniques based on this concept have been proposed in the literature. These techniques share the property that, in case of patterns having different lengths, a number of symbols are introduced in the shortest one, or deleted from the longest one, until both patterns have the same length. In case of applications in which strings are used for shape description, however, this property may introduce distortions in the shape, resulting in a distance measure not reflecting the perceived similarity between the shapes to compare. Moving from this consideration, we propose a new edit distance, called Weighted Edit Distance that does not require the introduction or the deletion of any symbol. Preliminary experiments performed by comparing our technique with the Normalized Edit Distance and the Markov Edit Distance have shown very encouraging results.

In this paper we propose an automatic updating system for urban vector maps that is able to detect changes between the old dataset (consisting of both vector and raster maps) and the present time situation represented in a raster map. In order to automatically detect as much changes as possible and to extract vector data for new buildings we present a system composed of three main parts: the first part detects changes between the input vector map and the new raster map (based on edge matching), the second part locates new objects (based on color segmentation), and the third part extracts new objects boundaries to be used for updating the vector map (based on edge detection, color segmentation and adaptive edge linking). Experiments on real datasets illustrate the approach.

In this paper, the problem of the surveillance and the security of indoor environments is addressed through the development of an autonomous surveillance vehicle (ASV). The ASV has been designed to perform the object detection by adopting an image alignment method followed by a change detection operation. Hence, in addition to the classical robotic tasks (e.g., navigation and obstacle avoiding), the tracking of objects (e.g., persons) moving in an indoor environment is considered. The tracking procedure allows the ASV to maintain the interesting objects in the centre of the image, and in specific cases to focus the image acquisition on particular parts of the object (e.g., face of a person, etc.) for recognition purposes. Experimental results have been performed on different real scenarios where no objects moves inside the monitored scene and where at least one moving object is into the scene.

In this paper we present a track matching algorithm based on the “major color” histograms matching and the post-matching integration useful for tracking a single object across multiple, limitedly disjoint cameras. First, the Major Color Spectrum Histogram (MCSH) is introduced to represent a moving object in a single frame by its most frequent colors only. Then, a two-directional similarity measurement based on the MCHS is used to measure the similarity of any two given moving objects in single frames. Finally, our track matching algorithm extends the single-frame matching along the objects’ tracks by a post-matching integration algorithm. Experimental results presented in this paper show the accuracy of the proposed track matching algorithm: the similarity of two tracks from the same moving objects has proved as high as 95%, while the similarity of two tracks from different moving objects has been kept as low as up to 28%. The post-matching integration step proves able to remove detailed errors occurring at the frame level, thus making track matching more robust and reliable.

This paper addresses the filtering problem when no assumption about linearity or gaussianity is made on the involved density functions. This approach, widely known as

particle filtering

, has been explored by several previous algorithms, including

Condensation.

Although it represented a new paradigm and promising results have been achieved, it has several unpleasant behaviours. We highlight these misbehaviours and propose an algorithm which deals with them. A test-bed, which allows proof-testing of new approaches, has been developed. The proposal has been successfully tested using both synthetic and real sequences.

In this paper a new approach to the problem of recognizing the preceding vehicle on highways is presented. The system is based on monocular vision. Since on highways the position of the preceding vehicle in the image varies slowly, its previous and current positions are compared using correlation. Such image processing produces a very clear output, which, at a higher level, allows a simple and fast recognition.

In this paper, it is assumed that each writer has his or her own statistics of handwriting displacement, therefore a statistical displacement analysis for handwriting verification is proposed. Here, a regularization method with the coarse-to-fine strategy computes the displacement function in questionable handwritten letters, and then it is normalized to remove the noisy displacement that arises from the position drift and scaling variation. Finally, the normalized displacement function and the statistics of displacement obtained in advance from registered authentic letters are used to calculate the distance from a standard handwritten letter to a questionable one. A fundamental simulation was conducted in order to evaluate the performance of the proposed method.

In this paper we describe a method for the reconstruction and visualization of functional models of monkey brains. Models are built through the registration of high resolution images obtained from the scanning of histological sections with reference photos taken during the brain slicing. From the histological sections it is also possible to acquire specifically activated neurons’ coordinates introducing functional information in the model. Due to the specific nature of the images (texture information is useless and the sections could be deformed when they were cut and placed on glass) we solved the registration problem by extracting corresponding cerebral cortex borders (extracted with a snake algorithm) and computing an image transform from the deformation linking them. The mapping is modeled as an affine deformation plus a non-linear field evaluated as an elastically constrained deformation minimizing contour distances. Registered images and contours are used then to build 3D models of specific brains by a software tool allowing the interactive visualization of cortical volumes together with the spatially referenced neurons classified and differently colored according to their functionalities.

In this paper we test the performance of two unsupervised clustering strategies for the analysis of LANDSAT multispectral images of the Temples of Paestum Area in Italy. The classification goal is to identify environmental factors (soils, vegetation types, water) on the images, exploiting the features of the seven LANDSAT spectral bands. The first strategy is a fast migrating means technique based on a Maximum Likelihood Principle (ISOCLUST algorithm),and the second is the Kohonen Self Organizing Map (SOM) neural network. The advantage of using the SOM algorithm is that both the information on classes and the similarity between the classes are obtained (since proximity corresponds to similarity among neurons). By exploiting the information on class similarity it was possible to automatically colour each cluster identified by the net (assigning a specific colour to each of them) thus facilitating a successive photo-interpretation.

This paper addresses some of the indications of the European Union for road safety by proposing a real-time traffic monitoring system for vehicle detection and tracking in bad illuminated scenarios. Several urban and extra-urban roads during the night or tunnels are characterized by low illumination, light spots, shadows, light reflections, etc. The main objectives of the proposed system are: (a) to monitor the traffic flow, (b) to estimate the vehicle’s speed or determine the state of the traffic, (c) to detect anomalous situations, e.g. rising alarms in case of road accidents or stopped cars. Experimental results on real image sequences demonstrate the effectiveness of the proposed system.

In this paper, we present a new approach to multi-camera object tracking based on the consistent labeling. An automatic and reliable procedure allows to obtain the homographic transformation between two overlapped views, without any manual calibration of the cameras. Object’s positions are matched by using the homography when the object is firstly detected in one of the two views. The approach has been tested also in the case of simultaneous transitions and in the case in which people are detected as a group during the transition. Promising results are reported over a real setup of overlapped cameras.