Computer Analysis of Images and Patterns

The ISO MPEG-7 Standard, also known as a Multimedia Content Description Interface, will be soon finalized. After several years of intensive work on technology development, implementation and testing by almost all major players in the digital multimedia arena, the results of this international project will be assessed by the most cruel and demanding judge: the market. Will it meet all the high expectations of the developers and, above all, future users? Will it result in a revolution, evolution or will it just simply pass unnoticed?In this invited lecture, I will review the components of the MPEG-7 Standard in the context of some novel applications. I will go beyond the classical image/video retrieval scenarios, and look into more generic image/object recognition framework relying on the MPEG-7 technology. Such a framework is applicable to a wide range of new applications. The benefits of using standardized technology, over other state-of-the art techniques from computer vision, image processing, and database retrieval, will be investigated. Demonstrations of the generic object recognition system will be presented, followed by some other examples of emerging applications made possible by the Standard. In conclusion, I will assess the potential impact of this new standard on emerging services, products and future technology developments.

This paper gives a brief introduction into the Visual part of the forthcoming new standard MPEG-7, the “Multimedia Content Description Interface”. It then emphasizes on the aspects how the Visual Descriptors of MPEG-7 were optimized for efficiency, compactness and behavior similar to human visual characteristics. The MPEG-7 descriptors were mainly designed for signal identification and recognition in the context of multimedia applications; however, they are applicable wherever interoperability between distributed systems designed for the task of visual information recognition need a standardized interface. I this sense, MPEG-7 may become a key element in the process of convergence of multimedia related applications with computer vision systems.

Color is one of the most important and easily identifiable features for describing visual content. The MPEG standardization group developed a number of descriptors that cover different aspects of this important visual feature. The objective of this paper is to review them and present some applications of the color descriptors by themselves and in combination with other visual descriptors.

We present three descriptors of texture feature of a region. Namely, the homogeneous texture descriptor (HTD), the edge histogram descriptor (EHD), and the perceptual browsing descriptor (PBD). They are currently included in the Committee Draft of the MPEG-7 Visual (ISO/TEC 15938-3). Each descriptor has a unique functionality and application domain. HTD and EHD describe statistical distribution of the texture and are useful for image retrieval application, while HTD is for homogeneously textured region and EHD is for multi-textured natural image or sketch. PBD is a compact descriptor suitable for quick browsing application.

We present an overview of the MPEG-7 motion descriptors viz. motion trajectory, camera motion, parametric motion and motion activity. These descriptors cover a wide range of functionality and hence enable several applications. We present the salient parts of the syntax, the semantics and the associated extraction of each of these descriptors. We discuss the possible applications for these descriptors and associated complexity trade-offs. We then describe a case study of a low complexity video browsing and indexing system that capitalizes on the simple extraction, compactness and effectiveness of the motion activity descriptor. This system relies on feature extraction in the compressed domain, which makes dynamic feature extraction possible. It combines the MPEG-7 motion activity descriptor and a simple color histogram to achieve both video summarization (top-down traversal) and indexing (bottom-up traversal) and thus enables a user-friendly video-browsing interface.

In this paper, we present the tools specified by the MDS part of the MPEG-7 standard for describing multimedia data such as images and video. In particular, we focus on the description tools that represent the structure and semantics of multimedia data to whose development we have actively contributed. We also describe some of our research prototype systems dealing with the extraction and application of MPEG-7 structural and semantic descriptions. These systems are AMOS, a video object segmentation and retrieval system, and IMKA, an intelligent multimedia knowledge application using the MediaNet knowledge representation framework.

This paper presents a very efficient and accurate method for retrieving images based on spatial color information. The method is based on a regular subblock approach with a large number of blocks and minimal color information for each block. Binary Thresholded Histogram and Extended Binary Thresholded Histogram are defined. Only 40 numbers are used to describe an image. Computing the distance is done by a very fast bitewise sum mod 2 operation.

Reversible transforms applied in wavelet coder to realize lossy-to-lossless compression are considered in this paper. 1-D wavelet transform possible to be customized in part of nowadays JPEG2000 standard is optimized to increase an efficiency of the first lossy phase of compression process. Different classes of reversible transforms were analyzed, evaluated in experiments, and compared one another in a sense of effectiveness, complexity and possibility of further optimization. Suitable selection of reversible wavelet transform can increase effectiveness of the coder even up to 2.7 dB of PSNR for 0.5 bpp in comparison to standard 5/3 transform. New reversible transform generated with lifting scheme was proposed. It overcomes all other in both phases of lossy-to-lossless compression (up to 0.4 dB of PSNR in comparison to the state-of-art transforms of JPEG2000 standardization process). Therefore, an efficiency of reversible wavelets can be comparable to irreversible wavelets effectiveness in several cases of lossy compression.

In the paper the new transform adapting to shapes of irregular image segments is introduced, the shape-adaptive (SA) DFT. Its definition is based on periodic data extension rather than data shifts, hence, in contrast to SA DCT segment reconstruction is possible even if part of contour data is missing. Visually the quality of images reconstructed from the part of SA DFT samples is almost as good as for the SA DCT, especially for high compression ratios.

Pairwise nearest neighbor method (PNN), in its exact form, provides good quality codebooks for vector quantization but at the cost of high run time. We consider the utilization of the partial distortion search technique in order to reduce the workload caused by the distance calculations in the PNN. By experiments, we show that the simple improvement reduces the run time down to 50–60%

The paper comprises study on accumulation of errors produced by near-lossless JPEG-LS in the consecutive compression-decompression cycles. Paper proves that alternatively, lossless compression can be performed on luminance and chrominance with reduced representation bit numbers. The advantage is that the errors do not accumulate in the consecutive cycles of compression because rounding errors of the RGB → YC_RC_B →RGB transformation do not accumulate in the consecutive transformation cycles. Exemplary experimental data that verify these statements are included in the paper.

Methods of lossless compression of medical image data are considered in this paper. Chosen classes of efficient algorithms were constructed, examined and optimised to conclude the most useful tools for creation of medical image representation. 2-D context-based prediction schemes, and statistical models of entropy coder were fitted to different characteristics of US, MR and CT images. The SSM technique of suitable-to-image characteristics scanning followed by statistical modelling of the context in arithmetic coder was found out as the most effective in most cases. Average bit rate value over test images is equal to 2.54 bpp for SSM coder and significantly overcomes 2.92 bpp achieved for CALIC. Efficient hybrid encoding method (SHEC) was proposed as a complex tool for medical image archiving and transmission. SHEC develops SSM by including CALIC-like coder for archiving the highest quality images and JPEG2000-like wavelet coder for transmission of high and middle quality images in telemedicine systems.

The problem of storage or transmission of codevectors is an essential issue in vector quantization with custom codebook. The proposed technique for compression of codebooks relies on structuring and ordering properties of a binary split algorithm used for codebook design. A simple algorithm is presented for automatic ordering of the codebook entries in order to group similar codevectors. This similarity is exploited in efficient compression of the codebook content by the means of lossless differential coding and lossy DCT-based coding. Experimental results of two compression experiments are reported and show that a small compression gain can be achieved in this way.

The paper describes a three-layer video coder based on spatiotemporal scalability and data partitioning. The base layer represents video sequences with reduced spatial and temporal resolution. Decoding of a middle layer gives full resolution images but with lower quality as compared to those obtained from the enhancement layer also. The bitrate overhead measured relative to the single layer MPEG-2 bitstream varies about 5% – 25% for progressive television test sequences. The base layer is fully MPEG-2 compatible and the whole structure exhibits high level of compatibility with individual building blocks of MPEG-2 coders. The paper reports experimental results that prove useful properties of the coder proposed.

A low bit rate video coding requires strict buffer regulations and low buffer delay. Thus a macroblock level rate control is necessary. However, an MB-level rate control is costly at low bit rates since there is an additional overhead if the quantization parameter is changed frequently within a frame. This paper presents the rate distortion model that selects the number of significant coefficients to code in a macroblock. To do this we derive the model for rate and distortion in terms of the number of encoded coefficients with pre-computed quantization accuracy. Rate-Distortion trade-off is solved then by the Lagrange optimization and the formula is obtained that indicate how many coefficients should be coded.

This article refers to the shape-adaptive DCT ( SA-DCT ) algorithm developed by Sikora and Makai in 1995. It is an important tool for encoding texture of arbitrary shaped video objects and can be included in MPEG-4 video codecs. In this paper a modification of normalized version SA-DCT redesigned for intraframe coding is presented. Simulations results show that this solution outperforms standard SA-DCT in rate-distortion sense. Efficiency is close to improved version SA-DCT for intraframe coding, known as ΔDC-SA-DCT. But computational overhead is smaller than for ΔDC-SA-DCT. Therefore, this solution may be attractive for hardware implementations.

This paper proposes a technique for object recognition using superquadric built models. Superquadrics, which are three dimensional models suitable for part-level representation of objects, are reconstructed from range images using the recover-and-select paradigm. Using an interpretation tree, the presence of an object in the scene from the model database can be hypothesized. These hypotheses are verified by projecting and re-fitting the object model to the range image which at the same time enables a better localization of the object in the scene.

This paper describes a new eigendecomposition method for weighted graph-matching. Although elegant by means of its matrix representation, the eigendecomposition method is proved notoriously susceptible to differences in the size of the graphs under consideration. In this paper we demonstrate how the method can be rendered robust to structural differences by adopting a hierarchical approach. We place the weighted graph matching problem in a probabilistic setting in which the correspondences between pairwise clusters can be used to constrain the individual correspondences. By assigning nodes to pairwise relational clusters, we compute within-cluster and between-cluster adjacency matrices. The modal co-efficients for these adjacency matrices are used to compute cluster correspondence and cluster-conditional correspondence probabilities. A sensitivity study on synthetic point-sets reveals that the method is considerably more robust than the conventional method to clutter or point-set contamination.

This paper investigates whether meaningful shape categories can be identified in an unsupervised way by clustering shock-trees. We commence by computing weighted and unweighted edit distances between shock-trees extracted from the Hamilton-Jacobi skeleton of 2D binary shapes. Next we use an EM-like algorithm to locate pairwise clusters in the pattern of edit-distances. We show that when the tree edit distance is weighted using the geometry of the skeleton, then the clustering method returns meaningful shape categories.

Genetic algorithms with a novel encoding scheme for feature selection are introduced. The proposed genetic algorithm is restricted to a particular predetermined feature subset size where the local optimal set of features is searched for. The encoding scheme limits the length of the individual to the specified subset size, whereby each gene has a value in the range from 1 to the total number of available features.This article also gives a comparative study of suboptimal feature selection methods using real-world data. The validation of the optimized results shows that the true feature subset size is significantly smaller than the global optimum found by the optimization algorithms.

This paper contributes to the statistical pattern recognition problem in which two classes of objects are considered and either of them is described by a mixture of Gaussian distributions. The components of either mixture are known, and unknown are only their weights. The class (state) of the object k is to be found at the mentioned incomplete a priori knowledge of the statistical model and the known observation x. The task can be expressed as a statistical decision making with non-random interventions. The task was formulated and solved first by Anderson and Bahadur [1] for a simpler case where each of two classes is described by a single Gaussian. The more general formulation with more Gaussians describing each of two classes was suggested by M.I. Schlesinger under the name generalized Anderson’s task (abbreviated GAT in the sequel). The linear solution to GAT was proposed in [5] and described recently in a more general context in a monograph [4].This contribution provides (i) a formulation of GAT, (ii) a taxonomy of various solutions to GAT including their brief description, (iii) the novel improvement to one of its solutions by proposing better direction vector for next iteration, (iv) points to our implementation of GAT in a more general Statistical Pattern Recognition Toolbox (in MATLAB, public domain) and (v) shows experimentally the performance of the improvement (iii).

The theoretical aspects of the decision rules stability problem are considered in the article. The new metric theorems of the stability of the polynomial decision rules are proven. These theorems are sequent from the well-known results of approximating irrationalities by rational numbers obtained by Liouville, Roth and Khinchin. The problem of optimal correlation between deterministic and stochastic methods and quality criterion in pattern recognition problems is also discussed.

The contribution is devoted to the recognition of objects and patterns deformed by imaging geometry as well as by unknown blurring. We introduce a new class of features invariant simultaneously to blurring with a centrosymmetric PSF and to affine transformation. As we prove in the contribution, they can be constructed by combining affine moment invariants and blur invariants derived earlier. Combined invariants allow to recognize objects in the degraded scene without any restoration.

This paper describes a technique for text segmentation of machine printed Gurmukhi script documents. Research in the field of segmentation of Gurmukhi script faces major problems mainly related to the unique characteristics of the script like connectivity of characters on the headline, two or more characters in a word having intersecting minimum bounding rectangles, multi-component characters, touching characters which are present even in clean documents. The segmentation problems unique to the Gurmukhi script such as horizontally overlapping text segments and touching characters in various zonal positions in a word have been discussed in detail and a solution has been proposed.

We describe an efficient algorithm for a calculation of 3D body volume and surface moments. The algorithm is based on implicit formulae for moment calculation and takes advantages of a polygonal representation. It uses only coordinates of the body vertices and facets orientation.

This paper investigates the use of evolutionary programming for the search of hypothesis space in visual learning tasks. The general goal of the project is to elaborate human-competitive procedures for pattern discrimination by means of learning based on the training data (set of images). In particular, the topic addressed here is the comparison between the ‘standard’ genetic programming (as defined by Koza [13]) and the genetic programming extended by local optimization of solutions, so-called genetic local search. The hypothesis formulated in the paper is that genetic local search provides better solutions (i.e. classifiers with higher predictive accuracy) than the genetic search without that extension. This supposition was positively verified in an extensive comparative experiment of visual learning concerning the recognition of handwritten characters.

Regenerating forest is important to account for carbon sink. Mapping regenerating forest from satellite data is difficult because it is spectrally mixed with natural forest. This paper investigated the combined use of texture features and voting classifications to enhance recognition of these two classes. Bagging and boosting were applied on Learning Vector Quantization (LVQ) and decision tree. Our results show that spatial textures improved separability. After applying voting classifications, class accuracy of decision tree increased by 5–7% and that of LVQ by approximately 3%. Substantial reduction (between 23% to 40%) of confusions between regenerating forest and natural forest were recorded. Comparatively, bagging is more consistent than boosting. An interesting observation is that even LVQ, a stable learner, was able to benefit from both voting classification algorithms.

This paper describes a novel technique for texture feature extraction and classification. The proposed feature extraction technique uses an Auto-Associative Neural Network (AANN) and the classification technique uses a Multi-Layer Perceptron (MLP) with a single hidden layer. The two approaches such as AANN-MLP and statistical-MLP were investigated. The performance of the proposed techniques was evaluated on large benchmark database of texture patterns. The results are very promising compared to other techniques. Some of the experimental results are presented in this paper.

Biomedical Engineering in present form started its developing since the late 1960’s and includes engineering applications in physiology and medicine, such as Biomechanics, Biomedical Instrumentation, Bioelectrical processes, Biocontrol systems, Biomedical signal and image processing, Medical informatics and others. In last decades Medical Imaging (MI) started to play important role in innovatory solutions and applications of biomedical engineering. In our presentation current trends of medical imaging development are considered. We mean an interesting projects, the ideas currently developed in labs and many research centers. Underlying our research leaded in many areas of medical imaging, nuclear and medical engineering, in collaborations with several medical and biomedical centers and institutes of physics and nuclear science, we intended to present a quick review of the most hopeful research directions. -What is important, and worth of work with? -Is the medical imaging dynamically developing science of the useful applications, truly important in an information society development, able to cumulate the resources and interests of youth? Subjectively, we tried to find the answers considering the following topics: functional imaging of organs and tissues: PET (brain), SPECT (circulatory system, organs), MRI (brain, circulatory system, organs), CT(circulatory system, organs); dynamic imaging of heart and blood vessels, blood supply of liver and kidneys, etc., 2-D and even 3-D perfusion maps, statistical flow models and objective computable parameters required to be standardized (EBCT, dynamic MRI, even US Power Doopler);image detectors (PET, radiography, CT), detection systems (SPECT), detectors (scintillators), sensors with amorphous silicon and selenium in digital radiography, x-ray tubes with laser beam irradiation;virtual endoscopy (bronchoscophy, gastroscophy);telemedicine, means protocols, network switches, satellite connectors, and PACS, DICOM servers, indexed data basis, Hospital Information Systems, remote health care, interactive consultations, patient and familyeducation, structure of safety access, hierarchical exam evaluation, teleconferences, inspection and quality control, etc.;medical image compression, JPEG2000 and other encoding lossy and lossless techniques necessary for efficient data storing and progressive transmission,computer-aided diagnosis: the examples of improvements in digital mammography, ultrasound systems; image-guided surgery and therapy, multimodal systems (PET, MRI, CT), 3-D imaging (acquisition, reconstruction) for various medical systems;physics of medical imaging, image and signal processing, physiology and function from multidimensional images, visualization and display procedures, image perception, observer performance, image quality evaluation tests and technology assessment; and others.Because of such wide range of these image engineering applications it is very difficult to select the most important perspective research. Presented ones were chosen to show the important from our point of view proofs of MI support necessity in modern diagnosis and therapy. Therefore, more elements of MI should be included in medical education at the Universities and propagated by Society organizations. An important conclusions derived from our study depicts predicted sources of increasing industrial development of MI and a role of MI which is expected to play in a future hospital clinical service.

Interaction increases flexibility of segmentation but it leads to undesired behaviour of an algorithm if knowledge being requested is inappropriate. In region growing, this is the case for defining the homogeneity criterion as its specification depends also on image formation properties that are not known to the user. We developed a region growing algorithm that learns its homogeneity criterion automatically from characteristics of the region to be segmented. It produces results that are only little sensitive to the seed point location and it allows a segmentation of individual structures. The method was successfully tested on artificial images and on CT images.

This is a feasibility study of tumor motion tracking in images generated by radiotherapy treatment beam. The objective is to control the beam during free breathing so as to reduce the irradiated zone and thus preserve healthy tissues. Two algorithms were tested on portal images sequences. Optical flow estimation (standard Horn and Schunck’s algorithm), applied to images from a patient, gave poor results because of low contrast and absence of texture in these images. Target tracking algorithm (block-matching), tested on images of a phantom with a radio-opaque marker, gave satisfactory results: mean absolute error was less than 1 mm. Hence, tumor tracking in portal images is possible, provided that a marker can be implanted in tumor’s vicinity. For images without markers, further work is necessary to assess if the small amount of motion information contained in such images can be reliably exploited.

A scheme for cell extraction in color histological images based edge detection and thinning is considered. An algorithm for thinning of color images is proposed that is based on thinning of pseudo gray-scale image. To extract accurately gray-scale levels, we propose a new coordinate system for color representation: system PHS, where P is a vector of color distance, H is a hue (chromaticity), S is a relative saturation. This coordinate system allows one to take into account specifics of histological images. Comparison of image thinning in other coordinate color systems is given that shows the image thinning in PHS system produces a rather high-quality skeleton of the objects in a color image. The proposed algorithm was tested on the histological images.

We are presenting a software tool developed for the purpose of atherosclerotic plaque study in high resolution Magnetic Resonance Images. A new implementation of balloon-type active contour model used for segmentation and quantification of blood vessel lumen is described. Its originality resides in a dynamic scaling process which makes the influence of the balloon force independent of the current size of the contour. The contour can therefore be initialized by single point. Moreover, system matrix inversion is performed only once. Hence computational cost is strongly reduced. This model was validated in ex vivo vascular images from Watanabe heritable hyperlipidaemic rabbits. Automatic quantification results were compared to measurements performed by experts. Mean quantification error was smaller than average intra-observer variability.

A new method of image reconstruction for active, pulse thermography is presented. Based on experimental results the thermal model of the observed object is proposed. Studies on thermal transients basing on the FEM object model are presented. Examples of reconstructed images are presented and described for phantoms and for in-vivo measurements. Possible applications are discussed.

In this work we model human hip joint. We determine certain parameters of the hip joint and make relation to existing parametrization based on two-dimensional radiological pictures. Initial proposals as to usefulness of these parameters are made. We develop a Java 3-D tool that facilitates obtaining 3D based parameters.

Investigation of heart rate variability is of considerable interest in physiology, clinical medicine and drug development. HRV analysis requires accurate rhythm classification. The well-known ARGUS system defines the useful method of rhythm classification. The original set of features used in this method contains 4 parameters: QRS duration, QRS height, QRS offset and QRS area. Zhou at al. showed, that the spatial features: T wave amplitude in lead V2, QRS and T axes angles in frontal plane, and QRS-T spatial angle are of utmost value for diagnostic classification of ventricular conduction defects. We studied usefulness of spatial features instead of original ones in the ARGUS system classification method. The spatial features were computed using SPART method developed by the authors. Classification results for spatial and original features are similar and close to those obtained by the original ARGUS system. The study results confirm usefulness of spatial features for automatic rhythm analysis.

Automatic segmentation of MR images is a complex task, particularly for structures which are barely visible on MR. Hippocampus is one of such structures. We present an active contour based segmentation algorithm, suited to badly defined structures, and test it on 8 hippocampi. The basic algorithm principle could also be applied for object tracking on movie sequences. Algorithm initialisation consists of manual segmentation of some key images. We discuss and solve numerous problems: partially blurred or discontinuous object boundaries; low image contrasts and S/N ratios; multiple distracting edges, surrounding the correct object boundaries. The active contours’ inherent limitations were overcome by encoding a priori geometric information into the deformation algorithm. We present a geometry encoding algorithm, followed by specializations needed for hippocampus segmentation. We validate the algorithm by segmenting normal and atrophic hippocampi. We achieve volumetric errors in the same range as those of manual segmentation (±5%). We also evaluate the results by false positive/negative errors and relative amounts of volume agreements.

In image registration it is vital to perform matching of those points in a pair of images which actually match each other, and to postpone those which do not match. It is not always known in advance, however, which points have their counterparts, and where are they located. To overcome this, we propose to use the Hausdorff distance function modified by using a voting scheme as a fitting quality function. This known function performs very well in guiding the matching process and supports stable matches even for low quality data. It also makes it possible to speed up the algorithms in various ways. An application to accuracy assessment of oncological radiotherapy is presented. Low contrast of images used to perform this task makes this application a challenging test.

We propose a novel segmentation method for real time lip tracker initialisation which is based on a Gaussian mixture model of the pixel data. The model is built using the Predictive Validation technique advocated in [4]. In order to construct an accurate model in real time, we adopt a quasi-random image sampling technique based on Sobol sequences. We test the proposed method on a database of 145 images and demonstrate that its accuracy, even with a few number of samples, is satisfactory and significantly better than the segmentation obtained by k-means clustering. Moreover, the proposed method does not require the number of segments to be specified a priori.

Particle Image Velocimetry (PIV) is a popular approach to flow visualisation in hydro- and aerodynamic studies and applications. The fluid is seeded with particles that follow the flow and make it visible. Traditionally, correlation techniques have been used to estimate the displacements of the particles in a digital PIV sequence. In this paper, two efficient feature tracking algorithms are customised and applied to PIV. The algorithmic solutions of the application are described. Techniques for coherence filtering and interpolation of a velocity field are developed. Experimental results are given, demonstrating that the tracking algorithms offer Particle Image Velocimetry a good alternative to the existing techniques.

This paper presents algorithms to implement the estimation of motion, focusing on the finite element method as a framework for the development of techniques. The finite element approach has the advantages of a rigorous mathematical formulation, speed of reconstruction, conceptual simplicity and ease of implementation via well-established finite element procedures in comparison to finite volume or finite difference techniques. The finite element techniques are implemented as a triangular discretisation, and preliminary results are presented. An important advantage is the capacity to tackle problems in which non-uniform sampling of the image sequence is appropriate, which will be addressed in future work.

In this paper we present a colour constancy algorithm for real-time face tracking. It is based on a modification of the well known Grey World algorithm in order to use the redundant information available in an image sequence. In the experiments conducted it is clearly more robust to sudden illuminant colour changes than popular the rg-normalised algorithm.

We present a framework for fast (linear time) local estimation of optical flow in image sequences. Starting from the commonly used brightness constancy assumption, a simple differential technique is derived in a first step. Afterwards, this approach will be extended by the application of a nonlinear diffusion process to the flow field in order to reduce smoothing at motion boundaries. Due to the ill-posedness of the determination of optical flow from the related differential equations, a Wavelet-Galerkin projection method is applied to regularize and linearize the problem.

A mathematical model of the motion of a solid particle in oscillatory stream of a viscous liquid was set up and analytically solved for Reynolds number in the relative motion Re < 2.0 (Stokes flow) and Reynolds number characterising the liquid stream Re* < 2100 (laminar flow). A computer aided method based on video image processing was applied as an experimental tool to analyse the particle motion and verify the mathematical model.

Log-polar imaging is an important topic in space-variant active vision, and facilitates some visual tasks. Translation estimation, though essential for active tracking, is more difficult in (log-)polar coordinates. We propose here a novel, conceptually simple, effective, and efficient method for translational motion estimation. It is based on a gradient-based minimization procedure. Experimental results with log-polar images using a software-based log-polar remmapper are presented.

Many different methods for tracking humans were proposed in the past several years, yet surprisingly only a few authors examined the accuracy of the proposed systems. As the accuracy analysis is impossible without the well-defined ground truth, some kind of at least partially controlled environment is needed. Analysis of an athlete motion in sport match is well suited for that purpose, and it coincides with the need of the sport research community for accurate and reliable results of motion acquisition. This paper presents a development of a two-camera people tracker, incorporating two complementary tracking algorithms. The developed system is suited for simultaneously tracking several people on a large area of a handball court, using a sequence of 384-by-288 pixel images from fixed cameras. We also examine the level of accuracy that this kind of computer vision system setup is capable of.

Augmented reality requires the geometric registration of virtual or remote worlds with the visual stimulus of the user. This can be achieved by tracking the head pose of the user with respect to the reference coordinate system of virtual objects. If tracking is achieved with head-mounted cameras, registration is known in computer vision as pose estimation. Augmented reality is by definition a real-time problem, so we are interested only in bounded and short computational time. We propose a new linear algorithm for pose estimation. Our algorithm shows better performance than the linear algorithm of Quan and Lan [14] and is comparable to the non-predicted time iterative approach of Kumar and Hanson [8].

The basic methodologies used in animation are presented and their most significant problems connected with combining real and virtual worlds are recognised. It includes the creation of virtual object and description of its realistic movement. To perform these tasks an optical method based on fringe projection and marker tracking is applied. The spatial analysis of a projected fringe pattern delivers an information about shape of an object and its out-of-plane displacement. Additionally the analysis of positions of fiducial points at the object observed by two CCDs provides an on-line information about (x,y,z) co-ordinates of these points. Combining the information about object shape and 3D co-ordinates of fiducial points during the measurement enables to generate a virtual model of the object together with the description of its movement. The concept described above is experimentally tested and the exemplary results are presented. The further works to implement this technique are discussed.

In this paper we propose a method for the automated acquisition of 3D human models for real-time animation. The individual to be modelled is placed in a monochrome environment and captured simultaneously by a set of 16 calibrated cameras distributed on a metal support above and around the head. A number of image sets is taken from various postures. A binary volume model will then be reconstructed from each image set via a shape-from-silhouette approach. Based on the surface shape and a reliable 3D skeletonisation of the volume model, a parametric human body template is fitted to each captured posture independently. Finally, from the parameter sets obtained initially, one unique set of posture-invariant parameters and the corresponding multiple sets of posture-dependent parameters are estimated using iterative optimisation. The resulting model consists of a fully textured triangular surface mesh over a bone structure, ready to be used in real-time applications such as 3D video-conferencing or off-the-shelf multi-player games.

In this paper, we propose computer assisted visualization for manual landmarking of specific points on cephalometric radiographs. The signal to noise ratio of radiographs is very low, because of superimposing of anatomical structures, dissymetries or artefacts. On radiographs of children, the localization of cephalometric points presents a great inter-subject, and inter- and intra-expert varibility, which is considerably reduced by considering an adaptative coordinates space. This coordinates space allows us to obtain statistical landmarking of cephalometric points used to define regions of interest. Each region takes advantage of a specific image processing, to enhance local and particular features (bone or suture). An augmented reality image is presented to the human expert, to focus on main sutures and bones in a small region of interest. This method is applied to the nettlesome problem of the interpretation of cephalometric radiographs, and provides satisfying results according to a cephalometric expert.

There are thousands of old black and white movies that are the cultural heritage of nations. These films are quite often seriously degraded. This is a problem of significant importance especially in Poland, where most of cinematic heritage was damaged during and after World War II. There is a wide spectrum of defects of different kinds and various complexity, which is a serious challenge for image processing scientists. In this paper a systematic methodology for solving these difficult problems is proposed. It contains an analysis of most common defects and introduces their taxonomy. The most important part of the work is devoted to the detection and removal of degradations. For this purpose different tools of image processing are applied, especially based on mathematical morphology. Considering the diversity and complexity of the defects one can easily observe that there is no uniform methodology that could be successfully applied to all degradation types. Unfortunately it does not seem to be possible to detect and remove all of them completely automatically. Therefore the whole system for semi-automatic treatment (with limited human interaction) is proposed.

A non-uniform distribution of rolling pressure during steel lamination may produce flatness asymmetries in the steel sheets, causing a certain curvature on its edges. This deformation may cause stoppages in the rolling process, and damages in the machinery. A computer vision system for measuring this straightness defect is presented. This system shows the adaptation of well-known computer vision techniques to fit precision and real-time constraints. Some problems that arise during the implementation phase are also described, and the correspondent solutions outlined.

Deflection compensation of flexible boom structures in robot positioning is becoming an important part of machine automation. Positioning is usually done using tables containing the magnitude of the deflection with inverse kinematics solutions of a rigid manipulator. In this paper, a method for locating the tip of a flexible manipulator using machine vision and a method for positioning the tip of a flexible manipulator using neural networks are proposed. A machine vision system was used in the data collection phase to locate the boom tip and the collected data was used to train MLP-networks. The developed methods improve the accuracy of manipulator positioning, and it can be integrated in the control system of the manipulator. The methods have been tested in real-time laboratory environment, and the results were promising. During the testing, the locating and the positioning were noticed to function as required, yielding reliable results with sufficient computation times.

We report on pattern recognition algorithms in discriminant analysis, which were used on Laser Induced Breakdown Spectroscopy (LIBS) spectra (intensity of signal against wavelength) for metal identification and sorting purposes. In instances where accurate elemental concentrations are not needed, discriminant analysis can be applied, to compare and match spectra of “unknown“ samples to library spectra of calibration samples. This type of “qualitative“ pattern recognition analysis has been used here for material identification and sorting. Materials of different matrix materials (e.g. Al, Cu, Pb, Zn, vitrification glass, steels, etc.) could be identified with 100% certainty, using Principle Component Analysis and the Mahalanobis Distance algorithms. The limits within which the Mahalanobis Distance indicate a match status of Yes, Possible or No were investigated. The factors, which dictate these limits in LIBS analysis, were identified - (i) spectrum reproducibility and (ii) the sample-to-sample homogeneity. If correctly applied the combination of pattern recognition algorithms and LIBS provide a useful tool for remote and in-situ material identification problems, which are of a more “identify-and-sort” nature (for example those in the nuclear industry).

Professional film scanners acting in real time (24 frames per second) are still very expensive. In most cases using a slide scanner of medium resolution equipped with additional device for transporting a film reel would be a reasonable solution. The main problem, however is a lack of accurate positioning mechanism in such sort of scanners. Therefore the position of each frame could be to some extent accidental. If frames are scanned separately from each other and this process is performed for all the frames of a movie there is usually a significant jitter in this sequence. This paper presents an efficient and simple method of obtaining jitter-free sequence i.e. finding the precise cinematic frame location in a picture that is the output of the scanning process. The procedure consists of two steps: rough estimation and the fine one. During the rough step the borders of the frame can be detected based on finding area of maximal brightness. In the second step the displacements among frame backgrounds are calculated. Additionally in order to avoid the fixed background problem the local constant component is eliminated in the postprocessing phase. As a final result a jitter is removed almost completely.

The problems of irregular 3D-objects manufacturing preparation are considered. The irregular surfaces digitizing process by video system is offered. The mathematical models of video digitizing process and software basic stages for computer 3D-models creation are shown. The forming of the computer 3D-models from video image as one from main stages of irregular 3D-objects compact reverse manufacturing is considered. The video system assubsystem of Compact Reverse Engineering System (CRES) is offered.

The paper reports development of an efficient algorithm identifying junctions in contour images. The theoretical model and selected results are presented. The algorithm combines local operators and an inexpensive matching. The algorithm scans incoming contour images with a circular window, and two levels of detection are used: structural identification and geometrical localisation. At the first level, the prospective junctions are identified based on topological properties of the window’s content. At the second level, the prospective location of the junction’s centre is determined, and a projection operator calculates its 1D profile. The junction is eventually accepted if the profile matches the corresponding template profile. The software implementation has been intensively tested, but the hardware implementation is the ultimate objective. Currently, a development system that directly maps C programs into FPGA structures can be used for this purpose.

This paper deals with super-resolution, i.e. the reconstruction of a high resolution image from a sequence of low resolution noisy and possibly blurred images. We have developed an iterative procedure for minimizing a measure of discrepancy based on the Csiszàr’s I-divergence. One advantage of this algorithm is to provide a natural positivity constraint on the solution. We consider a block-based version to speed up convergence and we propose a computationally efficient implementation. We also introduce a temporal multiscale version of the algorithm, which proves to be more robust and stable. The algorithm requires the computation of the apparent motion in the image sequence. We use a robust multiresolution estimator of a 2D parametric motion model in order to keep computational efficiency.

We give an overview of the problems for global polyhedrization in 3D digital images and present a solution by using our results of flatness analysis. We take both analytical and combinatorial topological approaches to define our flatness measurement which enable us to measure the degree of flatness for each point on a discretized object surface.

The paper describes an application of morphological mosaic interpolation based on distance function calculation to computer-aided animations. The existing method was extended and generalised in such a way that it can interpolate between any two mosaics - not only between two mosaics with non-empty intersection as the original method does. The problem for the proper generation of interpolation function - disappearing and moving particles, has been solved in the proposed method. An example of animation is also presented. It shows the change of the borders in Central Europe after the World War II.

In this paper, a class of openings and closings is investigated using the notion of propagation criteria. The main goal in studying these transformations consists in eliminating some inconveniences of the morphological opening (closing) and the opening (closing) by reconstruction. The idea in building these new openings and closings comes from the notions of morphological filters by recontruction and levelings. Morphological filters by reconstruction extract, from an image, the connected components that are marked. The reconstruction process of the input image is achieved using geodesic dilation (or erosion) until stability. However, since thin connections exist, these filters reconstructs too much and sometimes it is impossible to eliminate some undesirable regions. Because of this inconvenience, propagation criteria must be introduced.

In this work we propose an algorithm for segmentation of the text and non-text parts of document image using multiscale feature vectors. We assume that the text and non-text parts have different textural properties. M-band wavelets are used as the feature extractors and the features give measures of local energies at different scales and orientations around each pixel of the M × M bandpass channel outputs. The resulting multiscale feature vectors are classified by an unsupervised clustering algorithm to achieve the required segmentation, assuming no a priori information regarding the font size, scanning resolution, type layout etc. of the document.

This paper discusses the problem of how to approximate the length of a parametric curve γ: [0, T] → ℝn from points q_i =γ(t_i), where the parameters t_i are not given. Of course, it is necessary to make some assumptions about the distribution of the t_i: in the present paper ε-uniformity. Our theoretical result concerns an algorithm which uses piecewise-quadratic interpolants. Experiments are conducted to show that our theoretical estimates are sharp, and that the assumption of ε-uniformity is needed. This work may be of interest in computer graphics, approximation and complexity theory, digital and computational geometry, and digital image processing.

In this paper we propose a new algorithm of noise reduction in color images. The new technique of image enhancement is capable of reducing impulsive and Gaussian noise and it outperforms the standard methods of noise reduction. In the paper a smoothing operator, based on a random walk model and on a fuzzy similarity measure between pixels connected by a digital self avoiding path is introduced. The efficiency of the proposed method was tested on the test color image using the objective image quality measures and the results show that the new method outperforms standard noise reduction algorithms.

The Wigner distribution of a two-dimensional image function has the form of a four-dimensional Fourier transform of a correlation product r(x₁x₂,χ₁,χ₂) with respect to the spatial-shift variables χ₁ and χ₂ The corresponding ambiguity function has the form of the inverse Fourier transform of rx₁,x₂,χ₁,χ₂ with respect to spatial variables x₁ and x₂. There exist dual definitions in the frequency domain (f₁,f₂,μ₁,μ₂), where μ₁,μ₂ are frequency-shift variables. The paper presents the properties of these distributions and describes applications for image analysis.

In this paper, we propose a Markov random field (MRF) image segmentation model which aims at combining color and texture features. The theoretical framework relies on Bayesian estimation associated with combinatorial optimization (Simulated Annealing). The segmentation is obtained by classifying the pixels into different pixel classes. These classes are represented by multi-variate Gaussian distributions. Thus, the only hypothesis about the nature of the features is that an additive white noise model is suitable to describe the feature values belonging to a given class. Herein, we use the perceptually uniform CIE-L*u*v* color values as color features and a set of Gabor filters as texture features. We provide experimental results that illustrate the performance of our method on both synthetic and natural color images. Due to the local nature of our MRF model, the algorithm can be highly parallelized.

In this paper, using hypergraph theory, we introduce an image model called Adaptive Image Neighborhood Hypergraph (AINH). From this model we propose a combinatorial definition of noisy data. A detection procedure is used to classify the hyperedges either as noisy or clean data. Similar to other techniques, the proposed algorithm uses an estimation procedure to remove the effects of the noise. Extensive simulations show that the proposed scheme consistently works well in suppressing of impulsive noise.

The response of spring triticale to the infection of heads with mixture of Fusarium culmorum and F. avenaceum isolates was investigated with the application of colour image analysis of grains. The results seem to suggest that there is a strong relationship between declining values of the two yield components: kernels weight per spike and one thousand kernels weight (TKW), and the values of colour components H S and I (hue, saturation, intensity); the presence of the relationship is confirmed by high values of simple correlation coefficients. The technique applied in the experiment makes it possible to diagnose a case of Fusarium infestation at a nearly 90% probability, the result that creates the basis for further research towards the elaboration of a completely automatic system of colour image analysis of grain.

A new filtering approach designed to eliminate impulsive noise in color images, while preserving fine image details is presented in this paper. The computational complexity of the new filter is significantly lower than that of the Vector Median Filter. The comparison shows that the new filter outperforms the VMF, as well as other standard procedures used in color image processing, when the impulse noise is to be eliminated.

We consider the problem of image segmentation as the general image processing task. The proposed algorithm is modification of the approach described in [14]. Based on the multidimensional morphological filter theory a universal segmentation algorithm is developed. We also present the results of the described segmentation method on several examples containing grayscale images of different objects.

We report on a study of images with imperfectly ordered motifs (distorted 2D lattices). Our aim was to study the Fourier transform properties of the images according to the Fourier transform of perfect lattices and to improve methods for analysis of continuously distorted 2D lattices and techniques for 2D reconstruction. For this purpose, the locally normalized correlation is used in the method of correlation averaging. In order to reduce the image distortion the function of the distortion is approximated step by step by the Lagrange interpolation polynomials. The modification of the methods is demonstrated on electron micrographs of the S-layers of differently prepared Cyanobacteria.

This paper is a continuation to the work in [1], in which a new algorithm for skeletonization is introduced. The algorithm given there and implemented for script and text is applied here on images like pictures, medical organs and signatures. This is very important for a lot of applications in pattern recognition, like, for example, data compression, transmission or saving. Some interesting results have been obtained and presented in this article. Comparing our results with others we can conclude that if it comes to thinning of scripts, words or sentences our method is as good as some of the latest approaches, when considering cursive script. However, when it comes to pictures, signatures or other more complicated images, our algorithm showed better and more precise results [6].

We present a video segmentation algorithm that accurately finds object boundaries, and does not require any user assistance. After filtering the input video, markers are selected. Around each marker, a volume is grown by evaluating the local color and texture features. The grown volumes are refined and motion trajectories are extracted. Self-descriptors for each volume, mutual-descriptors for a pair of volumes are computed from trajectories. These descriptors designed to capture motion, shape as well as spatial characteristics of volumes. In the fine-to-coarse clustering stage, volumes are merged into objects by evaluating their descriptors. Clustering is carried out until the motion similarity of merged objects at that iteration becomes small. A multi-resolution object tree that gives the video object planes for every possible number of objects is generated. Test results prove the effectiveness of the algorithm.

In the first part of the paper a new theoretical approach to the problem of image segmentation is described. A method for automatic segmenting of an unknown number and unknown location of objects in an image has been proposed. This method is based on both local properties of neighbouring pixels and global image features. To allow for automated segmentation, slices are formed at different values of the threshold level, which contain spatial uniformity regions. In the second part, the image segmentation is considered as a problem of selection of slices, which should comprise regions with features satisfying the requirements desired. The selection is based on the proposed minima criterion including a volume analysis of neighbouring slices. An important characteristic of the approach is that it reflects object shapes devoid of noise, and does not use heuristic parameters such as an edge value. The results of this method are presented on several examples containing greyscale images of objects of different brightness.

A new method for extracting topographic features from images approximated by triangular meshes is presented. Peaks, pits, passes, ridges, valleys, and flat regions are defined by considering the topological and geometric relationship between the triangular elements. The approach is suitable for several computer-based recognition tasks, such as navigation of autonomous vehicles, planetary exploration, and reverse engineering. The method has been applied to a wide range of images, producing very promising results.

The ”seeded region growing” (SRG) is a segmentation technique which performs an image segmentation with respect to a set of initial points, known as seeds. Given a set of seeds, SRG then grows the regions around each seed, based on the conventional region growing postulate of similarity of pixels within regions. The choice of the seeds is considered as one of the key steps on which the performance of the SRG technique depends. Thus, numerous knowledge-based and pure data-driven techniques have been already proposed to select these seeds. This paper studies the usefulness of visual attention in the seed selection process for performing color image segmentation. The purely data-driven visual attention model, considered in this paper, provides the required points of attention which are then used as seeds in a SRG segmentation algorithm using a color homogeneity criterion. A first part of this paper is devoted to the presentation of the multicue saliency-based visual attention model, which detects the most salient parts of a given scene. A second part discusses the possibility of using the so far detected regions as seeds to achieve the region growing task. The last part is dedicated to experiments involving a variety of color images.

This paper analyzes four explicit, two implicit and four semi-implicit finite difference algorithms for the linear shape from shading problem. Comparisons of accuracy, solvability, stability and convergence of these schemes indicate that the weighted semi-implicit scheme and the box scheme are better than the other ones because they can be calculated more easily, they are more accurate, faster in convergence and unconditionally stable.

In this paper we present a new method for aligning point distribution models to noisy and unlabelled image data. The aim is to construct an enhanced version of the point distribution model of Cootes and Taylor in which the point-position information is augmented with a neighbourhood graph which represents the relational arrangement of the landmark points. We show how this augmented point distribution model can be matched to unlabelled point-sets which are subject to both additional clutter and point drop-out. The statistical framework adopted for this study interleaves the processes of finding point correspondences and estimating the alignment parameters of the point distribution model. The utility measure underpinning the work is the cross entropy between two probability distributions which respectively model alignment errors and correspondence errors. In the case of the point alignment process, we assume that the registration errors follow a Gaussian distribution. The correspondence errors are modelled using probability distribution which has been used for symbolic graph-matching. Experimental results are presented using medical image sequences.

This paper describes a new surface normal smoothing process which can be used in conjunction with shape-from-shading. Rather than directly smoothing the surface normal vectors, we exert control over their directions by smoothing the field of principal curvature vectors. To do this we develop a topography sensitive smoothing process which overcomes the problems of singularities in the field of principal curvature directions at the locations of umbilics and saddles. The method is evaluated on both synthetic and real world images.

The genetic algorithm is a simple optimization method for a wide variety of computer vision problems. However, its performance is often brittle and degrades drastically with increasing input problem complexity. While this problem is difficult to overcome due to the stochastic nature of the algorithm, this paper shows that a robust statistical design using repeated independent trials and hypothesis testing can be used to greatly alleviate the degradation. The working principle is as follows: The probability of success P of a stochastic algorithm A (genetic algorithm) can be estimated by running N copies of A simultaneously or running A repeatedly N times. By hypothesis testing, it is shown that P can be estimated to a required figure of merit (i.e. the level of significance). Knowing P, the overall probability of success P_repeated for N applications of A can be computed. This is used in turn to adjust N in an iterative scheme to maintain a constant P_repeated, achieving a robust feedback loop. Experimental results are reported on the application of this novel algorithm to an affine object detection problem.

A new approach to binocular stereo matching for epipolar geometry is presented. It is based on the idea that some features (edges) in the left image exert forces on similar features in the right image in order to attract them. Each feature point (i,j) of the right image is described by a coordinate x(i,j). The coordinates obey a system of time discrete Newtonian equations, which allow the recursive updating of the coordinates until they match the corresponding points in the left image. That model is very flexible. It allows shift, expansion and compression of image regions of the right image, and it takes into account occlusion to a certain amount. Furthermore, it can be implemented in parallel-sequential network structures allowing future real-time stereo processing (when corresponding hardware is available). The algorithm, which is confined here as a first step only to image points along edges, was applied to some stereo image pairs with a certain success, which gives hope for further improvements.

Recently multiple panoramic images have emerged and received increasingly interests in applications of 3D scene visualization and reconstruction. There is a need to characterize and clarify their common natures and differences so that a more general form/framework or a better computational model can be further discovered or developed. This paper introduces some notions at an abstract level for characterizing the essential components of panoramic image acquisition models. A general computational model is proposed to describe the family of cylindrical panoramas. The epipolar geometry of the cylindrical panoramic pairs for a general and a leveled case are particularly studied.

In many image classification problems the extent of usefulness of any variable for the purposes of discrimination apriori is unknown. This paper describes a unique fuzzy rule generation system developed to overcome this problem. By investigating interclass relationships very compact rule sets are produced with redundant variables removed. This approach to fuzzy system development is applied to two problems. The first is the classification of the Fisher Iris data [4] and the second is a road scene classification problem, based on features extracted from video images taken by a camera mounted in a motor vehicle.

This article introduces a segmentation method to automatically extract object parts from a reduced set of images. Given a database of objects and dividing all of them using local color histograms, we obtain an object part as the conjunction of the most similar ones. The similarity measure is obtained analyzing the behaviour of a local vector with respect to the whole object database. Furthermore, the proposed technique is able to associate an energy to each object part being possible to find the most discriminant object parts. We present the non-negative matrix factorization (NMF) technique to improve the internal data representation by compacting the original local histograms (50D instead of 512D). Moreover, the NMF based projected histograms only contain a few activated components and this fact improves the clustering results with respect to the use of the original local color histograms. We present a set of experimental results validating the use of the NMF in conjunction with the clustering technique.

The paper presents a comparison of a practical implementation of some area-based stereo algorithms. There are many stereo algorithms known that employ area based processing for matching of two or more images. However, much less information is available on practical implementations and applications of such methods, as well as their advantages and limitation. The work has been done to fill this gap and facilitate choice of the right stereo algorithm for machine vision applications, often using off-the-shelf cameras

A new autocalibration algorithm has been recently presented by Mendonça and Cipolla which is both simple and nearly globally convergent. Analysis of convergence is missing in the original article. This paper fills the gap, presenting an extensive experimental evaluation of the Mendonça and Cipolla algorithm, aimed at assessing both accuracy and sensitivity to initialization. Results show that its accuracy is fair, and - remarkably - it converges from almost everywhere. This is very significant, because most of the existing algorithms are either complicated or they need to be started very close to the solution.

An application-oriented vision-based traffic scene sensor system is designed. Its most important vision modules are identified and their algorithms are described in details: the on-line auto-calibration modules and three optional modules for 2-D measurement tasks (i.e. queue length detection, license plate identification and vehicle classification). It is shown that all three tasks may be regarded as applications of an iconic image classification scheme. Such a general scheme is developed and it can be applied for the above mentioned tasks by exchanging the application-dependent modules for pre-segmentation and feature extraction. The practical background of described work constitutes the IST project OMNI, dealing with the development of a network-wide intersection-driven model that can take advantage from the existence of advanced sensors, i.e. video sensors and vehicles equipped with GPS/GSM.

Active matching windows for matching in panoramic images taken by a catadioptric camera are proposed. The shape and the size of the windows vary depending on the position of an interest point. The windows size is then normalized and a standard correlation is used for measuring similarities of the points. A semi-iterative method based on sorting correspondences according to their similarity is suggested to remove possible outliers. It is experimentally shown that using this method the matching is successful for small and also big displacement of corresponding points.