Computer Vision and Graphics

The paper presents an architecture and features of a hybrid software/hardware system aimed at acceleration of image processing. Special design methodology of the software layers has been undertaken to meet the requirements of hardware and software cooperation. The front end consists of a fixed software interface. On the other hand, implementation variability is embedded into the complementary software and hardware layers, from which the best implementation can be selected in run time. The basic properties of the system are presented with some applications and discussion of further steps.

In spite of the fast progress of computer technologies there are many tasks, which need accelerated software methods to obtain the results in real time. We propose one of such accelerated logical-morphological methods to detect scratch-type objects on noisy surfaces. The method is based on the principles of human vision and includes an adaptive multithresholding and logical-morphological operations for the fusion of the object’s fragments.

This paper consider Spatio–Temporal Track–Before–Detect algorithm for tracking of moving objects using interlaced and progressive scan imaging sensors. Using the same image size and pixels’ readout rate interlaced and progressive scan approaches are compared. Especially the most important case of dim and pixel size target is analyzed due to spatial sampling theory violation for interlaced sensors applications.

The following paper proposes a procedure for an automatic detection of road-lanes, which involves three main steps: two-category scene segmentation, a detection of lane segment candidates and decision making. The main contribution of the paper is an introduction of a new step - high-curvature edge filtering - into a typical road image processing. This step substantially reduces an amount of noise and yields an improved lane detection performance. The remaining two steps of lane detection are dynamic programming-based selection of line segments, and Hough transform-based selection of the most consistent segments. The proposed approach proved to produce good results for a variety of real-traffic conditions, including correct detection in heavily shadowed road images.

In the paper a fast method of the digital image quality estimation is proposed. Our approach is based on the Monte Carlo method applied for some classical and modern full-reference image quality assessment methods, such as Structural Similarity and SVD-based measure. Obtained results are compared to the effects achieved using the full analysis techniques. Significant reduction of the number of analysed pixels or blocks leads to fast and efficient estimation of image quality especially in low performance systems where the processing speed is much more important than the accuracy of the quality assessment.

In the paper the analysis of the influence of the colour space on the results obtained during image quality assessment using the Structural Similarity index and the Singular Value Decomposition approach has been investigated. Obtained results have been compared to the ones achieved by widely used Normalised Colour Difference (NCD) metric. All the calculations have been performed using the LIVE Image Quality Assessment Database in order to compare the correlation of achieved results with the Differential Mean Opinion Score (DMOS) values obtained from the LIVE database. As a good solution for the further research, also with the use of some other image quality metrics, the application of the HSV colour space is proposed instead of commonly used YUV/YIQ luminance channel or the average of the RGB channels.

We propose a method of verification whether it is possible to measure image similarity by constructing a vector of metrics, regardless of what low-level features were extracted from images. We also present an on-line system which will be used to gather a dataset required to conduct the proposed experiment.

One important issue of the image compression is the quality of the resultant image. The Picture Quality Scale Measure (PQS) was used to improve this quality. This perceptual measure evaluates the visual distortions in color compressed image on the basis of the original and the resultant ones. This measure uses also a group of observers who are needed to evaluate the distortions. The implementation of PQS measure together with neural network enabled the independence of human observers. It was also possible to use this measure to influence the quality during the process of image compression. Each level of the wavelet transform was compressed with one wavelet function which caused the least visual distortions which were computed by Picture Quality Scale. The image was compressed with different transforms. The choice of a particular transform was taken on the ground of the PQS value. There were five different wavelet transforms to choose: Haar, Daubechies degree 4, Daubechies degree 6, Daubechies degree 8 and 5/3 transform. The image was coded with Embedded Zerotree Wavelet Coding (EZW).

This paper presents the method of evaluating the image quality using similarity of images phase spectrum. The authors introduce phase correlation coefficient as an objective measure of an image quality index which is compared to the subjective distortion evaluation using stimulus impairment scale. Artificially distorted images produced by proportional mixing of images phase spectra with a noise were used for testing purposes. The results are the mean correlation coefficient values related to the mean opinion score grades. The obtained relation between human responses and phase correlation is linear.

Creation of virtual reality models from photographs is very complex and time-consuming process, that requires special equipment like laser scanners, a large number of photographs and manual interaction. In this work we present a method for generating of surface geometry of photographed scene. Our approach is based on the phenomenon of shallow depth-of-field in close-up photography. Representing such surface details is useful to increase the visual realism in a range of application areas, especially biological structures or microorganisms.

For testing purposes a set of images of the same scene is taken from a typical digital camera with macro lenses with a different depth-of-field. Our new image fusion method employs discrete Fourier transform to designate sharp regions in this set of images, combine them together into a fully focused image and finally produce a height field map. Further image processing algorithms approximate three dimensional surface using this height field map and a fused image. Experimental results show that our method works for wide range of cases and gives a good tool for acquiring surfaces from a few photographs.

The paper discusses some aspects of the acquisition and preparation of 3D imaging data in two application areas: orthodontics and cultural heritage. Scanning devices used by our team are presented. Examples are given of the imaging of museum exhibits and other historic objects. Also described are possible modalities of presenting the 3D data in various hardware contexts and for various categories of users.

The paper presents a novel approach to the contour stitching method which improves the reconstruction of 3D objects from parallel contours. The innovation assumes the construction of outer and inner contour from many contours on one cross-section in case of branching. This exchange allows easy one to one connection, between contours on adjoining cross-sections, which has already been defined. Additionally, such a definition permits from artificial holes generation. Moreover, the method takes care of smooth surface reconstruction in case of concave and convex shapes. The performance of the novel algorithm has been tested on the set of artificial data as well as on the set of data gathered from CT scans and proved it to work well.

In the article we focused on

progressive meshes

which are the way to hierarchically represent 3D objects. We present the

Me3d

system developed to process and present digitalized objects of cultural heritage. The optimization of progressive meshes was described and the rendering performance of displaying the simplified meshes is analyzed.

Robust tracking of multi-objects is still challenging in real scenarios such as crowed scenes. In this paper a novel method in color image sequences is proposed for tracking multiple objects in non-cooperative situations. A system of independent particle filters with an adaptive motion model is used which tracks the moving objects under complex situations. Besides, in order to handle the conflicted situations, an integrated data association technique is exploited which adjusts the particle filters accordingly. Results have shown the good performance of the proposed method on various complex-situation image sequences.

Computer vision approaches to saliency are based, among others, on uniqueness [1], local complexity [2], distinctiveness [3,4], spectral variation [5], and irregularity [6]. Saliency can also be viewed as the information in the data relative to a representation or model [7]. When a representation is built, a residual error is often minimised. The residual can be used to obtain saliency maps for solving challenging tasks of image and video processing. We introduce the notion of the resonant SVD and demonstrate that the SVD residual at the resonant spacing is selective to defects in spatially periodic surface textures and events in time-periodic videos. Examples with real-world images and videos are shown and discussed.

This paper presents a new method for extracting pedestrian contours from images using 2D and 3D information obtained from a stereo-vision acquisition system. Two pedestrian contour types are extracted. First is obtained from static pedestrian confidence images using fixed background scenes and second from general traffic scenes having variable background. A robust approach for building contour hierarchies of these contours is then presented. First hierarchy is built of ”perfect” contours extracted from fixed background scenes and the second one is built of ”imperfect” contours extracted from images with variable background. The objective is to evaluate the two hierarchies in order to identify the best one for real time pedestrian detection.

Existing collision detection methods usually need long precalculation stage or difficult, time-consuming real-time computation. Moreover its effectiveness considerably decreases while the complexity of the scene or objects increases. Especially dynamic scenes with moving objects invoke necessity of each frame collisions recalculation due to changeable objects position. So far seemingly promising solutions supported by potential fields do not introduce satisfactory functionality as they are mainly devoted to static scenes with one predefined aim. This paper introduces method providing a new potential field construction which lets the camera reach both static and translating objects without constraints and protects user from getting into their structure while approaching volatile goals. Additionally proposed method comprises easy to bring through pre- calculation stage and becomes a scene-complexity independent solution.

The paper describes a simple quantitative model of AVC/ H.264 coders. The model defines the relationship between the bitstream and the quantization step (

Qstep

) for I- and P-frames. The whole allowed range of

Qstep

values has been divided into 3 intervals. In 1

st

and 3

rd

interval, the proposed model has only one parameter that depends on sequence content, whereas in 2nd interval the proposed model has three parameters that depend on sequence content. The experiments have been conducted on 4CIF sequences and showed that proposed model fits experimental data very well in all intervals.

The article presents a general outline of the signal theory using predict wavelet transform. The predict wavelet transform (1) is a new attitude to a multiresolution signal analysis by discrete wavelet transform [8]. It is implemented in accordance with a lifting scheme, which allows realizing a signal filtration by biorthogonal wavelets [10,9]. Thanks to it, there exists a possibility to receive an optimal, biorthogonal filter for analysis of chosen signal characteristic. The article describes the signal approximation method by lifting scheme (9). This method is a generalization of the predict interpolation proposed by Wim Sweldens [6,10]. It allows to predict odd signal samples using a polynomial degree much lower than an interpolation polynomial degree. This solution not only enables approximating by algebraic polynomial but also by some base functions. Thus, this method is more flexible and optimal. Using an orthogonal Gram’s polynomial (23) or a trigonometric polynomial for approximation eliminates a problem of ill conditioning of matrix coefficients (19). Such conditioning can cause large rounding errors during performing computer operation, so it leads to incorrect results. The article enables to get acquainted with the relation between classical and predict wavelet transform (6). The method of obtaining a biorthogonal highpass filter and lowpass filter (analysis and synthesis signal) is shown on the base of a second rank predictor example. The corresponding filters are received by summation of flow-ways between Predict and Update coefficients. The calculated filter coefficients are described by wavelet and scaling functions in a graphical form (7, 8). The summary chapter presents the use of lifting scheme during multiresolution image analysis (9) and irregular meshes analysis 3D (10).

The main goal of transcoding is to change bit rate of video sequence. This can be done by cascaded connection of decoder and encoder, known as Cascaded Pixel Domain Transcoder (CPDT). Decoding and re-encoding video bit stream always gives lower image quality than encoding original sequence. This paper presents a new technique of video transcoding that is able to deliver image quality superior to CPDT and has lower computational complexity. The technique is restricted to homogeneous (within the same bit stream format) transcoding of bit streams encoded according to H.264/AVC(MPEG 4) standard specification. The standard defines different types of encoded frames but proposed technique is designed for I(ntra) type frames only.

An approach for lossless and near-lossless compression of still images together with its system-level multi-core hardware model utilizing blending-prediction-based technique is presented in this paper. We provide a mathematical background of the proposed approach and utilize a Network on Chip type of connection in the hardware model which benefits from a new multi-path routing algorithm and heuristic algorithms for core mapping realizing subsequent stages of the compression algorithm. The experimental results confirming the advantages of the proposed approach are provided.

This paper presents a new robust and reliable marker less camera tracking system for outdoor augmented reality using only a mobile handheld camera. The proposed method is particularly efficient for partially known 3D scenes where only an incomplete 3D model of the outdoor environment is available. Indeed, the system combines an edge-based tracker with a sparse 3D reconstruction of the real-world environment to continually perform the camera tracking even if the model-based tracker fails. Experiments on real data were carried out and demonstrate the robustness of our approach to occlusions and scene changes.

This work proposes new static and dynamic based methods for facial expression recognition in stereo image sequences. Computer vision 3-d techniques are applied to determine real world geometric measures and to build a static geometric feature vector. Optical flow based motion detection is also carried out which delivers the dynamic flow feature vector. Support vector machine classification is used to recognize the expression using geometric feature vector while k-nearest neighbor classification is used for flow feature vector. The proposed method achieves robust feature detection and expression classification besides covering the in/out of plane head rotations and back and forth movements. Further, a wide range of human skin color is exploited in the training and the test samples.

This paper presents Local Rank Patterns (LRP) - novel features for rapid object detection in images which are based on existing features Local Rank Differences (LRD). The performance of the novel features is thoroughly tested on frontal face detection task and it is compared to the performance of the LRD and the traditionally used Haar-like features. The results show that the LRP surpass the LRD and the Haar-like features in the precision of detection and also in the average number of features needed for classification. Considering recent successful and efficient implementations of LRD on CPU, GPU and FPGA, the results suggest that LRP are good choice for object detection and that they could replace the Haar-like features in some applications in the future.

A common approach to pattern recognition and object detection is to use a statistical classifier. Widely used method is AdaBoost or its modifications which yields outstanding results in certain tasks like face detection. The aim of this work was to build real-time system for detection of dogs for surveillance purposes. The author of this paper thus explored the possibility that the AdaBoost based classifiers could be used for this task.

Meeting recording is being performed through microphones and video cameras in order to keep a permanent record of the events that are happening during the meetings. The technology to perform such recording is already mature and recording is being performed already for some time. However, efficient retrieval of the information from meeting data remains a hot topic of contemporary research. Several approaches to information retrieval exist, such as indexing the data, event semantics analysis of the data, etc. This contribution focuses on automatic video editing of the data in order to prepare audiovisual material, based on several audio and video sources, that is suitable for human users to see. The video editing takes the original audio and video data as its input as well as the results of analysis of audio and video streams and user instructions. The output of the method is a simple audiovisual stream.

In this paper we propose a video segmentation algorithm that in the final delineation of the object employs the graph-cut. A partitioning of the image based on pairwise region comparison is done at the beginning of each frame. A set of keypoints is tracked over time via optical flow to extract regions, which are likely to be parts of the object of interest. The tracked keypoints contribute towards better temporal coherence of the object segmentation. A probabilistic occupancy map of the object is extracted using such initial object segmentation and a probabilistic shape model. The map is utilized in a classifier that operates both on pixels and regions. The aim of the classifier is to extract a trimap consisting of foreground, background and unknown areas. The trimap is employed by graph-cut. The outcome of the graph-cut is used in on-line learning of the shape model. The performance of the algorithm is demonstrated on freely available test sequences.

In this paper we explore a multi-layer background change detection method based on projections of spatiotemporal 3D texture maps. The aim of this method is to provide a background change detection of a region viewed by multiple cameras. Camera views are projected onto a common ground plane, thus creating a spatially aligned multi-layer background. The aligned multi-layer background is subdivided into non-overlapping texture blocks, and block data is dimensionally reduced by principal component analysis. Motion detection is performed on each block, and non-moving sections of the block are clustered into multiple hyperspheres. An analysis of the clusters from spatially aligned multi-layer blocks reveal regions of changed background. This method is evaluated on surveillance videos available from PETS2006 and PETS2007 datasets.

Echocardiographic technology has currently reached a stage where it can provide 4D visual data revealing details of the real heart motion. Possibility of spatial reconstruction and quantitative description of such motion became very important task in today’s cardiology. Unfortunately, because of the low quality such image data does not allow precise measurements. To overcome this problem images need to be processed further and moving structures have to be extracted. In this work we present a method for estimating heart motion from 3D echocardiographic image sequence. We also introduce a novel method for quantitative and qualitative validation of motion reconstruction.

In the paper, the estimation algorithm of eye blinking for a computer animation system using biopotentials occurring on an actor’s face is considered. The measuring system incorporates minimal 3–electrodes configuration, which is relatively cheap solution. The proposed algorithm allows both the measurement of eyes orientation and detection of eyes blinking.

Computed understanding of CT images used for aided stroke diagnosis was the subject of reported research. Subtle hypodense changes of brain tissue as direct ischemia signs was estimated and extracted to improve diagnosis. Fundamental value of semantic content representation approximated from source images was studied. Nonlinear approximation of subtle pathology signatures in multiscale domain was verified for several local bases including wavelets, curvelets, contourlets and wedgelets. Different rationales for best bases selection were considered. Target pathology estimation procedures were optimized with a criterion of maximally clear extraction of diagnostic information. Visual expression of emphasized hypodenstity was verified for a test set of 25 acute stroke examinations. Suggested methods of stroke nonlinear approximation in many scales may facilitate the early CT-based diagnosis.

This paper present personal identification and verification based on iris pattern and propose a new algorithm for iris feature extraction. The algorithm is based on texture analysis using wavelet transform. Iris code is generated using representation of the wavelet coefficients based on a wedgelet dictionary. Preliminary results on CASIA iris image database confirm the effectiveness of this method and encourage further research.

This paper introduces a photon subsurface scattering method to simulate light transport in human colon tissue. First the theoretical model and parameters of human tissue including autofluorescence phenomenon was presented. Then it was described the Monte-Carlo model of steady-state light transport in multi-layered colon. The goal of this investigation is to simulate the light propagation in tissue and to collect the data containing the effect of fluorescence. This information will be used to generate images. Pictures taken for different adjustment of light parameters should define a configuration for which cancerous structures are visible quickly and precisely. Real medical devices can adjust their parameters to the simulated ones and help with efficient diagnosis and recognition of diseased structures.

Recognizing pedestrians in traffic scenarios is an important task for any smart vehicle application. Within the context of a real-time stereo based driving assistance system, this paper presents a novel method for recognizing pedestrians. We have designed a meta- classification scheme composed of a mixture of Bayesian and boosted classifiers that learn the discriminant features of a pedestrian space partitioned into attitudes like pedestrian standing and pedestrian running. Our experiments show that the mixture of classifiers proposed outperforms a single classifier trained on the whole un-partitioned object space. For classification we have used a probabilistic approach based on Bayesian Networks and Adaptive Boosting. Two types of features were extracted from the image: anisotropic gaussians and histograms of gradient orientations (HOG).

This paper addresses the issue of automatic wood defect classification. We propose a tree-structure support vector machine (SVM) to classify four types of wood knots by using images captured from lumber boards. Simple and effective features are proposed and extracted by first partitioning the knot images into 3 distinct areas, followed by applying an order statistic filter to yield an average pseudo color feature in each area. Excellent results have been obtained for the proposed SVM classifier that is trained by 800 wood knot images. Performance evaluation has shown that the proposed SVM classifier has resulted in an average classification rate of 96.5% and false alarm rate of 2.25% over 400 test knot images. Our future work includes more extensive tests on large data set and the extension of knot types.

One of the large problems for electrical power delivery through power lines in the Northern countries is when snow or ice accumulates on electrical insulators. This could lead to snow or ice-induced outages and voltage collapse, causing huge economic loss. This paper proposes a novel real-time automatic surveillance and image analysis system for detecting and estimating the snow and ice coverage on electric insulators using images captured from an outdoor 420 kV power transmission line. In addition, the swing angle of insulators is estimated, as large swing angles due to wind cause short circuits. Hybrid techniques by combining histogram, edges, boundaries and cross-correlations are employed for handling a broad range of scenarios caused by changing weather and lighting conditions. Experiments have been conducted on the captured images over several month periods. Results have shown that the proposed system has provided valuable estimation results. For image pixels related to snows on the insulator, the current system has yielded an average detection rate of 93% for good quality images, and 67.6% for images containing large amount of poor quality ones, and the corresponding average false alarm ranges from 9% to 18.1%. Further improvement may be achieved by using video-based analysis and improved camera settings.

A currently popular trend in object detection and pattern recognition is usage of statistical classifiers, namely AdaBoost and its modifications. The speed performance of these classifiers largely depends on the low level image features they are using: both on the amount of information the feature provides and the processor time of its evaluation. Local Rank Differences is an image feature that is alternative to commonly used haar wavelets. It is suitable for implementation in programmable (FPGA) or specialized (ASIC) hardware, but -as this paper shows -it performs very well on graphics hardware (GPU) used in general purpose manner (GPGPU, namely CUDA in this case) as well. The paper discusses the LRD features and their properties, describes an experimental implementation of the LRD in graphics hardware using CUDA, presents its empirical performance measures compared to alter native approaches, suggests several notes on practical usage of LRD and proposes directions for future work.

Developing fast and accurate 2D image processing algorithms is an important task for the practical use of cybernetics. This paper presents an algorithm for fast and accurate blob detection and extraction based on the usage of two parameters

ζ

and

χ

. The algorithm is aimed to work in the color domain to prevent any loss of information but can also be implemented on gray-scale images. Achieved regions of interest can be further processed to achieve high level description. The algorithm is implemented in Java environment in order to adduce results on different video devices and system platforms.

In this paper we compare the performance of face recognition systems based on two deformable shape models and on three classification approaches. Face contours have been extracted by using two methods: the Active Shapes and the Bayesian Tangent Shapes. The Normal Bayes Classifiers and the Minimum Distance Classifiers (based on the Euclidean and Mahalanobis metrics) have been designed and then compared w.r.t. the face recognition efficiency. The influence of the parameters of the shape extraction algorithms on the efficiency of classifiers has been investigated. The proposed classifiers have been tested both in the controlled conditions and as a part of the automatic face recognition system.

This paper proposes a hierarchical model for the recognition of deformable objects. Object categories are modelled by multiple views, views in turn consist of several parts, and parts consist of several features. The main advantage of the proposed model is that its nodes can be tuned with regard to the spatial selectivity. Every node in a category, views or part can thus take on the shape of a simple bag of features or a geometrically selective constellation model including all forms in between. Together with the explicit modelling of multiple views this allows for the modelling of categories with high intra-class variance. Experimental results show a high precision for the recognition of a character from a cartoon data base.

Animating virtual human-like characters has been a challenge in computer graphics for many years. Until now no single technique exists which addresses all of the arising problems. This paper presents a short overview of the methods and directions that were taken around the world to develop a solution to the subject. Following that, the paper concentrates on example-based algorithms and statistical analysis of the virtual character motion recorded using motion capture class devices or hand-made animations. The application of PCA technique on joint orientation correlations is introduced and two approaches to the crucial issue of quaternion linearization are compared with experimental results given for each of them. Then the summary of current state of research is provided and conclusions and future work possibilities are discussed.

The paper presents a novel approach to particle systems dedicated to the use in real-time 3D graphics. A new particle system architecture is presented, with hierarchical structure of objects, implementing a scene graph- specific interfaces which automatically incorporate them into the main graph of the whole interactive scene. With the uniform parameter passing scheme, the objects can be easily edited in a way similar to any other objects (like models, lights, sounds etc.), even during the simulation, with an instant visual effect. A complex hierarchy of objects grants a unique flexibility and extensibility; either by parameterization of existing objects in an editor or by implementing additional task-specific objects, which are automatically incorporated into the whole hierarchy, any specific behaviour of particles may be achieved, which greatly increases the level of visual attractiveness of generated effects.

In this paper we present solutions to most important deficiencies of the current version of OpenGL library which are connected to the lack of proper debugging facilities, state management inefficiency and bind-to-change object management scheme. Our solutions differ from the ones that have surfaced earlier in that, they are relatively lightweight, low-level and portable.

In the paper spatial component estimation for Track–Before–Detect (TBD) based motion capture systems is presented. Using TBD algorithms is possible to track markers at low Signal–to–Noise Ratio level that is a typical case in motion capture system. In the paper is considered special and the most popular rectangular capture area (cuboid volumen) for three configurations of image processing schemes: full frame processing, single camera optimized and multiple cameras optimized. In the article separate TBD processing for every camera is assumed.

The paper describes a method for the analysis of the content of a binary image in order to find its structure. The class of images it deals with consists of images showing a groups of objects connected one to another forming a graph-like structure. Proposed method extracts automatically this structure from image bitmap and produces graph adjacency matrix describing it. The method is based on morphological image processing, skeletonization and labelling.

We present a specialized visualization system dedicated to support analysis of dynamical processes in large spiking neural networks. The key features of the considered system are: clarity of visual representation, easy accessibility of multiple views and parameters related to the network analysis, advanced and flexible GUI and the system interoperability. In this paper we focus specifically on the implementation issues related to the network 3D representation, design of the graphical objects, visualisation functions and system performance.

In this paper a graph transformation using the parallel derivation approach is used to model the process of distribution and adaptation for computer aided design. It is based on earlier research in formal language theory, especially graph grammars, and distributed models. The motivation for the ideas presented here is given and some possible ways of application are described. The application of this idea by the graph distribution toolkit proposed as a multi-agent framework is also considered. The approach is illustrated by an example from the domain of flat layout design.

There are a variety of applications areas that take advantage of the availability of three-dimensional data sets. These objects are represented as complex polygonal surfaces formed by hundreds of thousands of polygons, which causes a significant increase in the cost of storage, transmission and visualisation. Such models are usually not only geometrically complex, but they may also have various surface properties such as colour, textures and temperature, etc. This paper presents a extension of lifting scheme for the multiresolution decomposition and reconstruction of irregular triangle surface meshes with additional attributes.