Image Analysis and Processing – ICIAP 2011

We formulate the problem of high order structural matching by applying

dominant cluster analysis

(DCA) to a direct product hypergraph (DPH). For brevity we refer to the resulting algorithm as DPH-DCA. The DPH-DCA can be considered as an extension of the game theoretic algorithms presented in [8] from clustering to matching, and also as a reduced version of reduced version of the method of ensembles of affinity relations presented in [6]. The starting point for our method is to construct a

K

-uniform direct product hypergraph for the two sets of higher-order features to be matched. Each vertex in the direct product hypergraph represents a potential correspondence and the weight on each hyperedge represents the agreement between two

K

-tuples drawn from the two feature sets. Vertices representing correct assignment tend to form a strongly intra-connected cluster, i.e. a dominant cluster. We evaluate the association of each vertex belonging to the dominant cluster by maximizing an objective function which maintains the

K

-tuple agreements. The potential correspondences with nonzero association weights are more likely to belong to the dominant cluster than the remaining zero-weighted ones. They are thus selected as correct matchings subject to the one-to-one correspondence constraint. Furthermore, we present a route to improving the matching accuracy by invoking prior knowledge. An experimental evaluation shows that our method outperforms the state-of-the-art high order structural matching methods[10][3].

The aim of this article is to introduce a computationally tractable mathematical model of the relation between the complex wavelet coefficients of two different images of the same scene. Because the two images are acquisitioned at distinct times, from distinct viewpoints, or by distinct sensors, the relation between the wavelet coefficients is far too complex to handle it in a deterministic fashion. This is why we consider adequate and present a probabilistic model for this relation. We further integrate this probabilistic framework in the construction of a new image registration algorithm. This algorithm has subpixel accuracy, and is robust to noise and to a large class of local variations like changes in illumination and even occlusions. We empirically prove the properties of this algorithm using synthetic and real data.

We present a kernel based approach for image de-noising in the spatial domain. The crux of evaluation for the kernel weights is addressed by a Bayesian regression. This approach introduces an adaptive filter, well preserving edges and thin structures in the image. The hyper-parameters in the model as well as the predictive distribution functions are estimated through an efficient iterative scheme. We evaluate our method on common test images, contaminated by white Gaussian noise. Qualitative results show the capability of our method to smooth out the noise while preserving the edges and fine texture. Quantitative comparison with the celebrated total variation (TV) and several wavelet methods ranks our approach among state-of-the-art denoising algorithms. Further advantages of our method include the capability of direct and simple integration of the noise PDF into the de-noising framework. The suggested method is fully automatic and can equally be applied to other regression problems.

A color image segmentation technique which exploits a novel definition of rough fuzzy sets and the rough–fuzzy product operation is presented. The segmentation is performed by partitioning each block in multiple rough fuzzy sets that are used to build a lower and a upper histogram in the HSV color space. For each bin of the lower and upper histograms a measure, called

τ

index, is computed to find the best segmentation of the image. Experimental results show that the proposed method retains the structure of the color images leading to an effective segmentation.

We present two novel contributions to the problem of region classification in scenery/landscape images. The first is a model that incorporates local cues with global layout cues, following the statistical characteristics recently suggested in [1]. The observation that background regions in scenery images tend to horizontally span the image allows us to represent the contextual dependencies between background region labels with a simple graphical model, on which exact inference is possible. While background is traditionally classified using only local color and textural features, we show that using new layout cues significantly improves background region classification. Our second contribution addresses the problem of correct results being considered as errors in cases where the ground truth provides the structural class of a land region (e.g., mountain), while the classifier provides its coverage class (e.g., grass), or vice versa. We suggest an alternative labeling method that, while trained using ground truth that describes each region with one label, assigns both a structural and a coverage label for each land region in the validation set. By suggesting multiple labels, each describing a different aspect of the region, the method provides more information than that available in the ground truth.

Dynamic Contrast Enhanced Magnetic Resonance Imaging (DCE-MRI) could be helpful in screening high-risk women and in staging newly diagnosed breast cancer patients. Selection of suspicious regions of interest (ROIs) is a critical pre-processing step in DCE-MRI data evaluation. The aim of this work is to develop and evaluate a method for automatic selection of suspicious ROIs for breast DCE-MRI. The proposed algorithm includes three steps: (i) breast mask segmentation via intensity threshold estimation; (ii) morphological operations for hole-filling and leakage removal; (iii) suspicious ROIs extraction. The proposed approach has been evaluated, using adequate metrics, with respect to manual ROI selection performed, on ten patients, by an expert radiologist.

In this paper, we propose an approach based on the level set method for segmenting SAR (Synthetic Aperture Radar) images. In particular, the segmentation process presented consists in the evolution of an initial curve, including the interested region, until it reaches the boundary of the area to be extracted. The procedure proposed allows to obtain the same result of segmentation independently of the initial position of the curve. The results are shown on both synthetic and real images. The analyzed images are SAR PRI (Precise Images), acquired during the mission ERS2.

This paper presents a specific algorithm for foreground object extraction in complex scenes where the background varies unpredictably over time. The background and foreground models are first constructed by using an adaptive mixture of Gaussians in a joint spatio-color feature space. A dynamic decision framework, which is able to take advantages of the spatial coherency of object, is then introduced for classifying background/foreground pixels. The proposed method was tested on a dataset coming from a real surveillance system including different sensors installed on board a moving train. The experimental results show that the proposed algorithm is robust in the real complex scenarios.

Time-of-flight (TOF) cameras are primarily used for range estimation by illuminating the scene through a TOF infrared source. However, additional background sources of illumination of the scene are also captured in the measurement process. This paper uses radiometric modelling of the signals emitted from the camera and a Lambertian reflectance model to develop a shadow segmentation algorithm. The proposed model is robust and is experimentally verified using real data.

The Tucker model with orthogonality constraints (often referred to as the HOSVD) assumes decomposition of a multi-way array into a core tensor and orthogonal factor matrices corresponding to each mode. Nonnegative Tucker Decomposition (NTD) model imposes nonnegativity constraints onto both core tensor and factor matrices. In this paper, we discuss a mixed version of the models, i.e. where one factor matrix is orthogonal and the remaining factor matrices are nonnegative. Moreover, the nonnegative factor matrices are updated with the modified Barzilai-Borwein gradient projection method that belongs to a class of quasi-Newton methods. The discussed model is efficiently applied to supervised classification of facial images, hand-written digits, and spectrograms of musical instrument sounds.

We investigate the implementation of multi-label classification algorithms with a reject option, as a mean to reduce the time required to human annotators and to attain a higher classification accuracy on automatically classified samples than the one which can be obtained without a reject option. Based on a recently proposed model of manual annotation time, we identify two approaches to implement a reject option, related to the two main manual annotation methods: browsing and tagging. In this paper we focus on the approach suitable to tagging, which consists in withholding either all or none of the category assignments of a given sample. We develop classification reliability measures to decide whether rejecting or not a sample, aimed at maximising classification accuracy on non-rejected ones. We finally evaluate the trade-off between classification accuracy and rejection rate that can be attained by our method, on three benchmark data sets related to text categorisation and image annotation tasks.

Image categorization is a challenging problem when a label is provided for the entire training image only instead of the object region. To eliminate labeling ambiguity, image categorization and object localization should be performed simultaneously. Discriminative Multiple Instance Learning (MIL) can be used for this task by regarding each image as a bag and sub-windows in the image as instances. Learning a discriminative MI classifier requires an iterative solution. In each round, positive sub-windows for the next round should be selected. With standard approaches, selecting only one positive sub-window per positive bag may limit the search space for global optimum; meanwhile, selecting all temporal positive sub-windows may add noise into learning. We select a subset of sub-windows per positive bag to avoid those limitations. Spatial relations between sub-windows are used as clues for selection. Experimental results demonstrate that our approach outperforms previous discriminative MIL approaches and standard categorization approaches.

A common approach in many classification tasks consists in reducing the costs by turning as many errors as possible into rejects. This can be accomplished by introducing a reject rule which, working on the reliability of the decision, aims at increasing the performance of the classification system. When facing multiclass classification, Error Correcting Output Coding is a diffused and successful technique to implement a system by decomposing the original problem into a set of two class problems. The novelty in this paper is to consider different levels where the reject can be applied in the ECOC systems. A study for the behavior of such rules in terms of Error-Reject curves is also proposed and tested on several benchmark datasets.

Amplitude spectra of natural images look surprisingly alike. Their shape is governed by the famous 1/

f

power law. In this work we propose a novel low parameter model for describing these spectra. The

Sum-of-Superellipses

conserves their common falloff behavior while simultaneously capturing the dimensions of variation—concavity, isotropy, slope, main orientation—in a small set of meaningful illustrative parameters. We demonstrate its general usefulness in standard computer vision tasks like scene recognition and image compression.

In this paper we propose a novel approach to combine information form multiple high-dimensional feature spaces, which allows reducing the computational time required for image retrieval tasks. Each image is represented in a “(dis)similarity space”, where each component is computed in one of the low-level feature spaces as the (dis)similarity of the image from one reference image. This new representation allows the distances between images belonging to the same class being smaller than in the original feature spaces. In addition, it allows computing similarities between images by taking into account multiple characteristics of the images, and thus obtaining more accurate retrieval results. Reported results show that the proposed technique allows attaining good performances not only in terms of precision and recall, but also in terms of the execution time, if compared to techniques that combine retrieval results from different feature spaces.

Document analysis often starts with robust signatures, for instance for document lookup from low-quality photographs, or similarity analysis between scanned books. Signatures based on OCR typically work well, but require good quality OCR, which is not always available and can be very costly. In this paper we describe a novel scheme for extracting discrete signatures from document images. It operates on points that describe the position of words, typically the centroid. Each point is extracted using one of several techniques and assigned a signature based on its relation to the nearest neighbors. We will discuss the benefits of this approach, and demonstrate its application to multiple problems including fast image similarity calculation and document lookup.

In spite of many advantages, multi-modal biometric recognition systems are vulnerable to spoof attacks, which can decrease their level of security. Thus, it is fundamental to understand and analyse the effects of spoof attacks and propose new methods to design robust systems against them. To this aim, we are developing a method based on

simulating

the fake score distributions of individual matchers, to evaluate the relative robustness of different score fusion rules. We model the score distribution of fake traits by assuming it lies between the one of genuine and impostor scores, and parametrize it by a measure of the relative distance to the latter, named

attack strength

. Different values of the attack strength account for the many different factors which can affect the distribution of fake scores. In this paper we present preliminary results aimed at evaluating the capability of our model to approximate realistic fake score distributions. To this aim we use a data set made up of faces and fingerprints, including realistic spoof attacks traits.

Thermal faceprint has been paramount in the last years. Since we can handle with face recognition using images acquired in the infrared spectrum, an unique individual’s signature can be obtained through the blood vessels network of the face. In this work, we propose a novel framework for thermal faceprint extraction using a collection of graph-based techniques, which were never used to this task up to date. A robust method of thermal face segmentation is also presented. The experiments, which were conducted over the UND Collection C dataset, have showed promising results.

In this paper we present a method removing reflection of people on shiny floors in the context of people detection for video analysis applications. The method exploits chromatic properties of the reflections and does not require a geometric model of the objects. An experimental evaluation of the proposed method, performed on a significant database containing several publicly available videos, demonstrates its effectiveness. The proposed technique also favorably compares with respect to other state of the art algorithms for reflection removal.

The ability to predict, given an image or a video, where a human might fixate elements of a viewed scene has long been of interest in the vision community.

In this note we propose a different view of the gaze-shift mechanism as that of a motor system implementation of an active random sampling strategy that the Human Visual System has evolved in order to efficiently and effectively infer properties of the surrounding world. We show how it can be exploited to carry on an attentive analysis of dynamic scenes.

We propose a new simplified 3D body model (called SARC3D) for surveillance application, which can be created, updated and compared in real-time. People are detected and tracked in each calibrated camera, with their silhouette, appearance, position and orientation extracted and used to place, scale and orientate a 3D body model. For each vertex of the model a signature (color features, reliability and saliency) is computed from 2D appearance images and exploited for matching. This approach achieves robustness against partial occlusions, pose and viewpoint changes. The complete proposal and a full experimental evaluation are presented, using a new benchmark suite and the PETS2009 dataset.

We present a novel non-object centric approach for discovering activity patterns in dynamic scenes. We build on previous works on video scene understanding. We first compute simple visual cues and individuate elementary activities. Then we divide the video into clips, compute clip histograms and cluster them to discover spatio-temporal patterns. A recently proposed clustering algorithm, which uses as objective function the Earth Mover’s Distance (EMD), is adopted. In this way the similarity among elementary activities is taken into account. This paper presents three crucial improvements with respect to previous works: (i) we consider a variant of EMD with a robust ground distance, (ii) clips are represented with circular histograms and an optimal bin order, reflecting the atomic activities’similarity, is automatically computed, (iii) the temporal dynamics of elementary activities is considered when clustering clips. Experimental results on publicly available datasets show that our method compares favorably with state-of-the-art approaches.

This paper describes a method to detect and extract pedestrians trajectories in proximity of a sliding door access in order to automatically open the doors: if a pedestrian walks towards the door, the system opens the door. On the other hand if the pedestrian trajectory is parallel to the door, the system does not open. The sensor is able to self-adjust according to changes in weather conditions and environment. The robustness of this system is provided by a new method for disparity image extraction.

The rationale behind this work is that the device developed in this paper avoids unwanted openings in order to decrease needs for maintenance, and increase building efficiency in terms of temperature (i.e. heating and air conditioning). The algorithm has been tested in real conditions to measure its capabilities and estimate its performance.

The automatic processing and estimation of view direction and head pose in interactive scenarios is an actively investigated research topic in the development of advanced human-computer or human-robot interfaces. Still, current state of the art approaches often make rigid assumptions concerning the scene illumination and viewing distance in order to achieve stable results. In addition, there is a lack of rigorous evaluation criteria to compare different computational vision approaches and to judge their flexibility. In this work, we make a step towards the employment of robust computational vision mechanisms to estimate the actor’s head pose and thus the direction of his focus of attention. We propose a domain specific mechanism based on learning to estimate stereo correspondences of image pairs. Furthermore, in order to facilitate the evaluation of computational vision results, we present a data generation framework capable of image synthesis under controlled pose conditions using an arbitrary camera setup with a free number of cameras. We show some computational results of our proposed mechanism as well as an evaluation based on the available reference data.

This paper presents a novel approach for the automatic generation of image slideshows. Counter to standard cross-fading, the idea is to operate the image transitions keeping the subject focused in the intermediate frames by automatically identifying him/her and preserving face and facial features alignment. This is done by using a novel Active Shape Model and time-series Image Registration. The final result is an aesthetically appealing slideshow which emphasizes the subject. The results have been evaluated with a users’ response survey. The outcomes show that the proposed slideshow concept is widely preferred by final users w.r.t. standard image transitions.

In this paper, we propose a new supervised Neighborhood Discriminative Manifold Projection (NDMP) method for feature extraction in video-based face recognition. The abundance of data in videos often result in highly nonlinear appearance manifolds. In order to extract good discriminative features, an optimal low-dimensional projection is learned from selected face exemplars by solving a constrained least-squares objective function based on both local neighborhood geometry and global manifold structure. The discriminative ability is enhanced through the use of intra-class and inter-class neighborhood information. Experimental results on standard video databases and comparisons with state-of-art methods demonstrate the capability of NDMP in achieving high recognition accuracy.

Over the past several decades, pattern classification based on subspace methodology is one of the most attractive research topics in the field of computer vision. In this paper, a novel probabilistic linear subspace approach is proposed, which utilizes hybrid way to capture multi-dimensional data extracting maximum discriminative information and circumventing small eigenvalues by minimizing statistical dependence between components. During features extraction process, local region is emphasized for crucial patterns representation, and also statistic technique is used to regularize these unreliable information for both reducing computational cost and maintaining accuracy purposes. Our approach is validated with a high degree of accuracy with various face applications using challenging databases containing different variations.

This paper describes a novel approach to the computation of refractive index from polarisation information. Specifically, we use the refractive index measurements to gauge the quality of fruits and vegetables. We commence by using the method of moments to estimate the components of the polarisation image computed from intensity images acquired by employing multiple polariser angles. The method uses photometric stereo to estimate surface normals and then uses the estimates of surface normal, zenith angle and polarisation measurements to estimate the refractive index. The method is applied to surface inspection problems. Experiments on fruits and vegetables at different stages of decay illustrate the utility of the method in assessing surface quality.

Indirect reflection component degrades the preciseness of 3-D measurement with structured light projection. In this paper, we propose a method to suppress the indirect reflection components by spatial synchronous detection of structured light modulated with MLS (Maximum Length Sequence, M-sequence). Our method exploits two properties of indirect components; one is the high spatial frequency component which is attenuated through the scattering of projected light, and the other is the geometric constraint between projected light and its corresponding pixel of camera.

Several experimental results of measuring translucent or concave objects show the advantage of our method.

In the last years, the 3D reconstruction of surfaces which represent objects photographed by simple digital cameras has become more and more necessary to the scientific community. Through the most various mathematical and engineering methods, scientists continue to study the Shape-from-shading problem, using the photometric stereo technique which allows the use of several light sources, but keeps the camera at the same point of view. Several studies, through different advances on the problem, have checked that in the applications, the smallest number of photos that have to be considered is three. In this article we analyze the possibility to determine the objects’ surface using two images only.

Planckian illuminants and von Kries diagonal model are commonly assumed by many computer vision algorithms for modeling the color variations between two images of a same scene captured under two different illuminants. Here we present a method to estimate a von Kries transform approximating a Planckian illuminant change and we show that the Planckian assumption constraints the von Kries coefficients to belong to a ruled surface, that depends on physical cues of the lights. Moreover, we provide an approximated parametric representation of such a surface, making evident the dependence of the von Kries transform on the light color temperature and on the intensity.

We present and evaluate a novel idea for scalable lossy colour image coding with Matching Pursuit (MP) performed in a transform domain. The idea is to exploit correlations in RGB colour space between image subbands after wavelet transformation rather than in the spatial domain. We propose a simple quantisation and coding scheme of colour MP decomposition based on Run Length Encoding (RLE) which can achieve comparable performance to JPEG 2000 even though the latter utilises careful data modelling at the coding stage. Thus, the obtained image representation has the potential to outperform JPEG 2000 with a more sophisticated coding algorithm.

Color line extraction is an important part of the segmentation process. The proposed method is the generalization of the Gradient Line Detector (GLD) to color images. The method relies on the computation of a color gradient field. Existing color gradient are not “oriented”: the gradient vector direction is defined up to

π

, and not up to 2

π

as it is for a grey-level image. An oriented color gradient which makes use of an ordering of colors is proposed. Although this ordering is arbitrary, the color gradient orientation changes from one to the other side of a line; this change is captured by the GLD. The oriented color gradient is derived from a generalization from scalar to vector: the components of the gradient are defined as a “signed” distance between weighted average colors, the sign being related to their respective order. An efficient averaging method inspired by the Gaussian gradient brings a scale parameter to the line detector. For the distance, the simplest choice is the Euclidean distance, but the best choice depends on the application. As for any feature extraction process, a post-processing is necessary: local maxima should be extracted and linked into curvilinear segments. Some preliminary results using the Euclidean distance are shown on a few images.

This paper presents a bottom-up approach for perceptual segmentation of natural images. The segmentation algorithm consists of two consecutive stages: firstly, the input image is partitioned into a set of blobs of uniform colour (pre-segmentation stage) and then, using a more complex distance which integrates edge and region descriptors, these blobs are hierarchically merged (perceptual grouping). Both stages are addressed using the Combinatorial Pyramid, a hierarchical structure which can correctly encode relationships among image regions at upper levels. Thus, unlike other methods, the topology of the image is preserved. The performance of the proposed approach has been initially evaluated with respect to ground-truth segmentation data using the Berkeley Segmentation Dataset and Benchmark. Although additional descriptors must be added to deal with textured surfaces, experimental results reveal that the proposed perceptual grouping provides satisfactory scores.

A main issue in the archaeological research is the identification of colored surfaces and soils through the application of Munsell system. This method widely used also in other fields, like geology and anthropology, is based on the subjective matching between the real color and its standardized version on Munsell chart. For preventing many possible errors caused by the subjectivity of the system itself, in this paper an automatic method of color detection on selected regions of digital images of archaeological pottery is presented.

The paper focuses on the possibilities of color image retrieval of the images sharing the similar location of particular color or set of colors present in the depicted scene. The main idea of the proposed solution is based on treating image as a multispectral object, where each of its spectral channels shows locations of pixels of 11 basis colors within the image. Thus, each of the analyzed images has associated signature, which is constructed on the basis of the mixture approximation of its spectral components. The ability of determining of highly similar images, in terms of one or more basic colors, reveals that the proposed method provides useful and efficient tool for robust to impulse distortions image retrieval.

In the paper the first stage of the approach for automatic identification of football players is presented. It is based on the numbers placed on their shirts. The method works with video frames extracted during a television sport broadcast. The element of the system described in this paper is devoted to the localisation of the numbers and their extraction for future recognition. It is simple, yet efficient and it is based on the use of appropriate ranges in various colour spaces. Four colour spaces were experimentally evaluated for this purpose. Thanks to this, the ranges could be established for particular kits. Firstly, the part of an image with a shirt was localised, and later, within this area, a number was found.

This paper presents a real-time hand gesture recognizer based on a color glove. The recognizer is formed by three modules. The first module, fed by the frame acquired by a webcam, identifies the hand image in the scene. The second module, a feature extractor, represents the image by a nine-dimensional feature vector. The third module, the classifier, is performed by means of

Learning Vector Quantization

. The recognizer, tested on a dataset of 907 hand gestures, has shown very high recognition rate.

Achieving a high fidelity triangle mesh from 3D digital reconstructions is still a challenge, mainly due to the harmful effects of outliers in the range data. In this work, we discuss these artifacts and suggest improvements for two widely used volumetric integration techniques: VRIP and Consensus Surfaces (CS). A novel contribution is a hybrid approach, named IMAGO Volumetric Integration Algorithm (IVIA), which combines strengths from both VRIP and CS while adds new ideas that greatly improve the detection and elimination of artifacts. We show that IVIA leads to superior results when applied in different scenarios. In addition, IVIA cooperates with the hole filling process, improving the overall quality of the generated 3D models. We also compare IVIA to Poisson Surface Reconstruction, a state-of-the-art method with good reconstruction results and high performance both in terms of memory usage and processing time.

The conventional approach to the detection of microcalcifications on mammographies is to employ a sliding window technique. This consists in applying a classifier function to all the subwindows contained in an image and taking each local maximum of the classifier as a possible position of a microcalcification. Although effective such an approach suffers from the high computational burden due to the huge number of subwindows contained in an image. The aim of this paper is to experimentally verify if such problem can be alleviated by a detection system which employs a cascade-based localization coupled with a clustering algorithm which exploits both the spatial coordinates of the localized regions and a confidence degree estimated on them by the final stage of the cascade. The first results obtained on a publicly available set of mammograms show that the method is promising and has large possibility of improvement.

In the framework of Palaeography, the use of digital image processing techniques has received increasing attention in recent years, resulting in a new research field commonly denoted as “digital palaeography”. In such a field, a key role is played by both pattern recognition and feature extraction methods, which provide quantitative arguments for supporting expert deductions. In this paper, we present a pattern recognition system which tries to solve a typical palaeographic problem: to distinguish the different scribes who have worked together to the transcription of a single medieval book. In the specific case of a high standardized book typology (the so called Latin “Giant Bible”), we wished to verify if the extraction of certain specifically devised features, concerning the layout of the page, allowed to obtain satisfactory results. To this aim, we have also performed a statistical analysis of the considered features in order to characterize their discriminant power. The experiments, performed on a large dataset of digital images from the so called “Avila Bible” - a giant Latin copy of the whole Bible produced during the XII century between Italy and Spain - confirmed the effectiveness of the proposed method.

We describe a methodology for exploring the parameter spaces of rigid-body registrations in 3-D. It serves as a tool for guiding and assisting a user in an interactive registration process. An exhaustive search is performed over all positions and rotations of the template, resulting in a 6-D volume, or fitness landscape. This is explored by the user, who selects and views suitable 3-D projections of the data, visualized using volume rendering. The 3-D projections demonstrated here are the maximum and average intensity projections of the rotation parameters and a projection of the rotation parameter for fixed translation parameters. This allows the user to jointly visualize projections of the parameter space, the local behaviour of the similarity score, and the corresponding registration of the two volumes in 3-D space for a chosen point in the parameter space. The procedure is intended to be used with haptic exploration and interaction. We demonstrate the methodology on a synthetic test case and on real molecular electron tomography data using normalized cross correlation as similarity score.

We consider a Multiple Kernel Learning (MKL) framework for nuclei classification in tissue microarray images of renal cell carcinoma. Several features are extracted from the automatically segmented nuclei and MKL is applied for classification. We compare our results with an incremental version of MKL, support vector machines with single kernel (SVM) and voting. We demonstrate that MKL inherently combines information from different input spaces and creates statistically significantly more accurate classifiers than SVMs and voting for renal cell carcinoma detection.

Nanotechnology tools, such as Atomic Force Microscopy (AFM), are now becoming widely used in life sciences and biomedicine. AFM is a versatile technique that allows studying at the nanoscale the morphological, dynamic, and mechanical properties of biological samples, such as living cells, biomolecules, and tissues in their native state under physiological conditions. In this article, an overview of the principles of AFM will be first presented and this will be followed by discussion of some of our own recent work on the applications of AFM imaging to biomedicine.

The high dimensionality of some real life signals makes the usage of the most common signal processing and pattern recognition methods unfeasible. For this reason, in literature a great deal of research work has been devoted to the development of algorithms performing dimensionality reduction. To this aim, an useful help could be provided by the estimation of the intrinsic dimensionality of a given dataset, that is the minimum number of parameters needed to capture, and describe, all the information carried by the data. Although many techniques have been proposed, most of them fail in case of noisy data or when the intrinsic dimensionality is too high. In this paper we propose a local intrinsic dimension estimator exploiting the statistical properties of data neighborhoods. The algorithm evaluation on both synthetic and real datasets, and the comparison with state of the art algorithms, proves that the proposed technique is promising.

In this paper we present a novel dynamic programming algorithm to synthesize an optimal decision tree from

OR

-decision tables, an extension of standard decision tables, which allow to choose between several alternative actions in the same rule. Experiments are reported, showing the computational time improvements over state of the art implementations of connected components labeling, using this modelling technique.

In image processing, convolution is a frequently used operation. It is an important tool for performing basic image enhancement as well as sophisticated analysis. Naturally, due to its necessity and still continually increasing size of processed image data there is a great demand for its efficient implementation. The fact is that the slowest algorithms (that cannot be practically used) implementing the convolution are capable of handling the data of arbitrary dimension and size. On the other hand, the fastest algorithms have huge memory requirements and hence impose image size limits. Regarding the convolution of huge images, which might be the subtask of some more sophisticated algorithm, fast and correct solution is essential. In this paper, we propose a fast algorithm implementing exact computation of the shift invariant convolution over huge multi-dimensional image data.

This paper presents an algorithm of half ellipse detection from color images. Additionally the algorithm detects two color average values along the both sides of a half ellipse. In contrast to standard methods the new one finds not only parameters of the entire ellipse but also the end points of a half ellipse. The paper introduces a new way of edge and line detection. The new detector of edges in color images was designed to extract color on the both sides of an edge. The new line detector is designed to optimize the detection of endpoints of a line.

The distribution of digital images with the classic and newest technologies available on Internet (e.g., emails, social networks, digital repositories) has induced a growing interest on systems able to protect the visual content against malicious manipulations that could be performed during their transmission. One of the main problems addressed in this context is the authentication of the image received in a communication. This task is usually performed by localizing the regions of the image which have been tampered. To this aim the received image should be first registered with the one at the sender by exploiting the information provided by a specific component of the forensic hash associated with the image. In this paper we propose a robust alignment method which makes use of an image hash component based on the Bag of Visual Words paradigm. The proposed signature is attached to the image before transmission and then analyzed at destination to recover the geometric transformations which have been applied to the received image. The estimator is based on a voting procedure in the parameter space of the geometric model used to recover the transformation occurred to the received image. Experiments show that the proposed approach obtains good margin in terms of performances with respect to state-of-the art methods.

The Focus of Expansion (FoE) sums up all the available information on translational ego-motion for monocular systems. It has also been shown to present interesting features in cognitive research. As such, its localization bears great importance, either for robotic applications, as well as for attention fixation research. It will be shown that the so-called C-Velocity framework can be inversed in order to extract the FoE position from a rough scene structure estimation. This method rely on robust cumulative framework and only exploit the optical flow field relative norm as such, it is robust to angular noise and bias on the absolute optical flow norm.

Analysis of the retinal photoreceptor mosaic can provide vital information in the assessment of retinal disease. However, visual analysis of photoreceptor cones can be both difficult and time consuming. The use of image processing techniques to automatically count and analyse these photoreceptor cones would be beneficial. This paper proposes the use of multi-scale modelling and normalized cross-correlation to identify retinal cones in image data obtained from a modified commercially available confocal scanning laser ophthalmoscope (CSLO). The paper also illustrates a process of synthetic data generation to create images similar to those obtained from the CSLO. Comparisons between synthetic and manually labelled images and the automated algorithm are also presented.

Traditionally feature extraction is focussed on edge and corner detection, however, more recently points of interest and blob like features have also become prominent in the field of computer vision and are typically used to determine correspondences between two images of the same scene. We present a new approach to a Hessian blob detector, designed within the finite element framework, which is similar to the multi-scale approach applied in the SURF detector. We present performance evaluation that demonstrates the accuracy of our approach in comparison to well known existing algorithms.

Support Vector Machines are excellent binary classifiers. In case of multi–class classification problems individual classifiers can be collected into a directed acyclic graph structure DAGSVM. Such structure implements One-Against-One strategy. In this strategy a split is created for each pair of classes, but, because of hierarchical structure, only a part of them is used in the single classification process.

The number of classifiers may be reduced if their classification tasks will be changed from separation of individual classes into separation of groups of classes. The proposed method is based on the similarity of classes. For near classes the structure of DAG stays immutable. For the distant classes more than one is separated with a single classifier. This solution reduces the classification cost. At the same time the recognition accuracy is not reduced in a significant way. Moreover, a number of SV, which influences on the learning time will not grow rapidly.

In this paper new error measures to evaluate image features in 3D scenes are proposed and reviewed. The proposed error measures are designed to take into account feature shapes, and ground truth data can be easily estimated. As other approaches, they are not error-free and a quantitative evaluation is given according to the number of wrong matches and mismatches in order to assess their validity.

The Bayesian approach applied to image denoising gives rise to a regularization problem. Total variation regularizers have been introduced with the motivation of being edge preserving. However we show here that this may not always be the best choice in images with low/medium frequency content like digital radiographs. We also draw the attention on the metric used to evaluate the distance between two images and how this can influence the choice of the regularization parameter. Lastly, we show that hyper-surface regularization parameter has little effect on the filtering quality.

Variations in pose, illumination and expression in faces make face recognition a difficult problem. Several researchers have shown that faces of the same individual, despite all these variations, lie on a complex manifold in a higher dimensional space. Several methods have been proposed to exploit this fact to build better recognition systems, but have not succeeded to a satisfactory extent. We propose a new method to model this higher dimensional manifold with available data, and use a reconstruction technique to approximate unavailable data points. The proposed method is tested on Sheffield (previously UMIST) database, Extended Yale Face database B and AT&T (previously ORL) database of faces. Our method outperforms other manifold based methods such as Nearest Manifold and other methods such as PCA, LDA Modular PCA, Generalized 2D PCA and super-resolution method for face recognition using nonlinear mappings on coherent features.

In this work we propose an efficient and original method for ellipse detection which relies on a recent contour representation based on arcs and line segments [1]. The first step of such a detection is to locate ellipse candidate with a grouping process exploiting geometric properties of adjacent arcs and lines. Then, for each ellipse candidate we extract a compact and significant representation defined from the segment and arc extremities together with the arc middle points. This representation allows then a fast ellipse detection by using a simple least square technique. Finally some first comparisons with other robust approaches are proposed.

We consider the problem of extracting the morphology of a terrain discretized as a triangle mesh. We discuss first how to transpose Morse theory to the discrete case in order to describe the morphology of triangulated terrains. We review algorithms for computing Morse decompositions, that we have adapted and implemented for triangulated terrains. We compare the the Morse decompositions produced by them, by considering two different metrics.

One of the most recurrent problem in digital image processing applications is segmentation. Segmentation is the separation of components in the image: the ability to identify and to separate objects from the background. Depending on the application, this activity can be very difficult and segmentation accuracy is crucial in order to obtain reliable results. In this paper we propose an approach for spot detection in images with noisy background. The overall approach can be divided in three main steps: image segmentation, region labeling and selection. Three segmentation algorithms, based on global or local thresholding technique, are developed and tested in a real-world petroleum geology industrial application. To assess algorithm accuracy we use a simple voting technique: by a visual comparison of the results, three domain experts vote for the best algorithms. Results are encouraging, in terms of accuracy and time reduction, especially for the algorithm based on local thresholding technique.

Research in automatic facial expression recognition has permitted the development of systems discriminating between the six prototypical expressions, i.e. anger, disgust, fear, happiness, sadness and surprise, in frontal video sequences. Achieving high recognition rate often implies high computational costs that are not compatible with real time applications on limited-resource platforms. In order to have high recognition rate as well as computational efficiency, we propose an automatic facial expression recognition system using a set of novel features inspired by statistical moments. Such descriptors, named as statistical-like moments extract high order statistic from texture descriptors such as local binary patterns. The approach has been successfully tested on the second edition of Cohn-Kanade database, showing a computational advantage and achieving a performance recognition rate comparable than methods based on different descriptors.

Self-update and co-update algorithms are aimed at gradually adapting biometric templates to the intra-class variations. These update techniques have been claimed to be effective in capturing variations occurring in medium time period but no experimental evaluations have been done in the literature to clearly show this fact. The aim of this paper is the analysis and comparison of these update techniques on the sequence of input batch of samples as available over time, specifically, in the time-span of 1.5 years. Effectiveness of these techniques have been compared in terms of capability to capture significant intra-class variations and the attained performance improvement, over time. Experiments are carried out on DIEE multi-modal dataset, explicitly collected for this aim. This dataset is publicly available by contacting the authors.

We present a biologically motivated approach to fast feature extraction on hexagonal pixel based images using the concept of eye tremor in combination with the use of the spiral architecture and convolution of non-overlapping Laplacian masks. We generate seven feature maps “a-trous” that can be combined into a single complete feature map, and we demonstrate that this approach is significantly faster than the use of conventional spiral convolution or the use of a neighbourhood address look-up table on hexagonal images.

Appearance description is a relevant field in computer vision that enables object recognition in domains as re-identification, retrieval and classification. Important cues to describe appearance are colors and textures. However, in real cases, texture detection is challenging due to occlusions and to deformations of the clothing while person’s pose changes. Moreover, in some cases, the processed images have a low resolution and methods at the state of the art for texture analysis are not appropriate.

In this paper, we deal with the problem of localizing real textures for clothing description purposes, such as stripes and/or complex patterns. Our method uses the entropy of primitive distribution to measure if a texture is present in a region and applies a quad-tree method for texture segmentation.

We performed experiments on a publicly available dataset and compared to a method at the state of the art[16]. Our experiments showed our method has satisfactory performance.

In this paper, we evaluate the effectiveness and efficiency of the global image descriptors and their distance metric functions in the domain of object recognition and near duplicate detection. Recently, the global descriptor GIST has been compared with the bag-of-words local image representation, and has achieved satisfying results. We compare different global descriptors in two famous datasets against mean average precision (MAP) measure. The results show that Fuzzy Color and Texture Histogram (FCTH) is outperforming GIST and several MPEG-7 descriptors by a large margin. We apply different distance metrics to global features so as to see how the similarity measures can affect the retrieval performance. In order to achieve the goal of lower memory cost and shorter retrieval time, we use the Spectral Hashing algorithm to embed the FCTH in the hamming space. Querying an image, from 1.26 million images database, takes 0.16 second on a common notebook computer without losing much searching accuracy.

For most HVS(Human Visual System) perceptual models, the JND(Just Noticeable Difference) values in highly-textured image regions have little difference with those in edge areas. This is not consistent with the characteristics of human vision. In this paper, an improved method is introduced to give a better content-based perceptual mask than traditional ones using the arrangement of scale-invariant feature points. It could decrease the JND values in edge areas of those traditional masks so that they have an obvious difference with values in highly textured areas. Experimental results show the advantages of this improved approach visually, and the enhancement of the invisibility of watermarks.

In this paper we propose a new method to detect cropped images by analyzing the blocking artefacts produced by a previous block based compression techniques such as JPEG and MPEG family that are the most used compression standards for still images and video sequences. It is useful for image forgery detection, in particular when an image has been cropped. The proposed solution is very fast compared to the previous art and the experimental results show that it is quite reliable also when the compression ratio is low, i.e. the blocking artefact is not visible.

In this paper we present a dense 3D reconstruction pipeline from monocular video sequences using jointly Photometric Stereo (PS) and Structure from Motion (SfM) approaches. The input videos are completely uncalibrated both from the multi-view geometry and photometric stereo aspects. In particular we make use of the 3D metric information computed with SfM from a set of 2D landmarks in order to solve for the bas-relief ambiguity which is intrinsic from dense PS surface estimation. The algorithm is evaluated over the CMU Multi-Pie database which contains the images of 337 subjects viewed under different lighting conditions and showing various facial expressions.

Normalized convolution techniques operate on very few samples of a given digital signal and add missing information, trough spatial interpolation. From a practical viewpoint, they make use of data really available and approximate the assumed values of the missing information. The quality of the final result is generally better than that obtained by traditional filling methods as, for example, bilinear or bicubic interpolations. Usually, the position of the samples is assumed to be random and due to transmission errors of the signal. Vice versa, we want to apply normalized convolution to compress data. In this case, we need to arrange a higher density of samples in proximity of zones which contain details, with respect to less significant, uniform parts of the image. This paper describes an evolutionary approach to evaluate the position of certain samples, in order to reconstruct better images, according to a subjective definition of visual quality. An extensive analysis on real data was carried out to verify the correctness of the proposed methodology.

Shape-from-shading is an interesting approach to the problem of finding the shape of a face because it only requires one image and no subject participation. However, SfS is not accurate enough to produce good shape models. Previously, SfS has been combined with shape models to produce realistic face reconstructions. In this work, we aim to improve the quality of such models by exploiting a probabilistic SfS model based on Fisher-Bingham 8-parameter distributions (FB

8

). The benefits are two-fold; firstly we can correctly weight the contributions of the data and model where the surface normals are uncertain, and secondly we can locate areas of shadow and facial hair using inconsistencies between the data and model. We sample the FB

8

distributions using a Gibbs sampling algorithm. These are then modelled as Gaussian distributions on the surface tangent plane defined by the model. The shape model provides a second Gaussian distribution describing the likely configurations of the model; these distributions are combined on the tangent plane of the directional sphere to give the most probable surface normal directions for all pixels. The Fisher criterion is used to locate inconsistencies between the two distributions and smoothing is used to deal with outliers originating in the shadowed and specular regions. A surface height model is then used to recover surface heights from surface normals. The combined approach shows improved results over the case when only surface normals from shape-from-shading are used.

In this paper we introduce a new method for Visual Saliency detection. The goal of our method is to emphasize regions that show rare visual aspects in comparison with those showing frequent ones. We propose a bottom up approach that performs a new technique based on low level image features (texture) analysis. More precisely, we use SIFT Density Maps (SDM), to study the distribution of keypoints into the image with different scales of observation, and its relationship with real fixation points. The hypothesis is that the image regions that show a larger distance from the mode (most frequent value) of the keypoints distribution over all the image are the same that better capture our visual attention. Results have been compared to two other low-level approaches and a supervised method.

Digital image processing techniques are increasingly applied to the study of cultural heritage, namely to the analysis of paintings and archaeological artefacts, in order to identify particular features or patterns and to improve the readability of the artistic work. Digital or ‘virtual’ restoration provides a useful framework where comparisons can be made and hypotheses of reconstruction proposed without action or damage for the original object, according to the adopted general rules for practical restoration.