Computer Vision and Graphics

A large number of cameras record video around the clock, producing huge volumes. Processing these huge chunks of videos demands plenty of resources like time, man power, and hardware storage etc. Video summarization plays an important role in this context. It helps in efficient storage, quick browsing, and retrieval of large collection of video data without losing important aspects. In this paper, we categorize video summariztion methods on the basis of methodology used, provide detailed description of leading methods in each category, and discuss their advantages and disadvantages. Moreover, we discuss the situation in which each method is most suitable to use. The advantage of this research is that one can quickly learn different video summarization techniques, and select the method that is the most suitable according to one’s requirements.

Modeling of soft pelvic organs and their thicknesses is a difficult task, especially when inputs are noisy and scattered. In order to define the geometric step for a global pelvic surgery simulator, we define a new method based only on geometry while considering the problem of error transfer between outer and inner organ surfaces. We compare this approach with a parametric formulation and a mass-spring system.

In this paper an approach based on hybrid, evolutionary-neural computations to the IFS inverse problem is presented. Having a bitmap image we look for an IFS having the attractor approximating of a given image with a good accuracy. A method using IFSes consisting of a variable number of mappings is proposed. A genom has hierarchical structure. A number of different operators acting on various levels of the genome are introduced. The algorithm described in [7] is aided by multi-layer neural networks. Such improved algorithm is less time consuming.

In this paper we present a new method to reconstruction of images with additive Gaussian noise. In order to solve this inverse problem we use stochastic differential equations with reflecting boundary (in short reflected SDEs). The continuous model of the image denoising is expressed in terms of such equations. The reconstruction algorithm is based on Euler’s approximations of solutions to reflected SDEs.

We consider a classical Euler scheme with random terminal time and controlled parameter of diffusion. The reconstruction time of our method is substantially reduced in comparison with classical Euler’s scheme. Our numerical experiments show that the new algorithm gives very good results and compares favourably with other image denoising filters.

In this paper we explore the problem of reconstruction of vector-valued images with additive Gaussian noise. In order to solve this problem we use backward stochastic differential equations. Our numerical experiments show that the new approach gives very good results and compares favourably with deterministic partial differential equation methods.

Color quantization is an important operation with many applications in graphics and image processing. Clustering methods based on the competitive learning paradigm, in particular self-organizing maps, have been extensively applied to this problem. In this paper, we investigate the performance of the batch neural gas algorithm as a color quantizer. In contrast to self-organizing maps, this competitive learning algorithm does not impose a fixed topology and is insensitive to initialization. Experiments on publicly available test images demonstrate that, when initialized by a deterministic preclustering method, the batch neural gas algorithm outperforms some of the most popular quantizers in the literature.

The paper presents a collaborative project that offers stand-alone software applications for end-users and a complete open-source platform to rapidly develop/prototype medical image processing work-flows with sophisticated visualization and user interactions. It builds on top of a flexible cross-platform framework (Linux, Windows and MacOS) developed in

C++

, which guarantees an easy connection of heterogeneous

C++

modules and provides the user with libraries of high-level components to construct graphical user interfaces (GUI) including input/output (file management), display, interaction, data processing, etc.

In this article, we illustrate the usefulness of this framework through a research project dealing with the study of thrombosis in intra-cranial aneurysms. Algorithms developed by the researchers, such as image segmentation, stent model generation, its interactive virtual deployment in the segmented vessels, as well as the generation of meshes necessary to simulate the blood flow through thus stented vessels, have been implemented in a user-friendly GUI with 3D visualization and interaction.

The aim of this paper is to present a new way of multidimensional data visualization for explorative forecast built for real meteorological data coming from the Institute of Meteorology and Water Management (IMGW) in Katowice, Poland. In the earlier works two first authors of the paper proposed a method that aggregates huge amount of data based on fuzzy numbers. Explorative forecast uses similarity of data describing situations in the past to those in the future. 2D and 3D visualizations of multidimensional data can be used to carry out its analysis to find hidden information that is not visible in the raw data e.g. intervals of fuzziness, fitting real number to a fuzzy number.

Amorphous shapes have always been a challenge to CG modelers. Apparently natural in look, their topology is hard to retrieve manually. Scientists from different backgrounds have tried to understand and model such phenomena. Fractals belong to the most representative solutions, but still are rare in 3D domain. Methods for generation of fractal objects use mainly quaternion representations for nonlinear systems in four dimensions. In such a case advanced volumetric graphics methods need to be applied to convey multidimensional information. In the paper, we propose a simple and effective approach to use four-dimensional escape-time fractals as automated shape generator. We extend general quadratic fractal maps to four dimensions. The algorithm results in diverse aesthetically balanced volumetric shapes delivered in real time on a modern PC.

Architecture of algorithmically optimized MPEG-4 AVC/ H.264 video encoder is presented in the paper. The paper reveals details of implementation for the proposed MPEG-4 AVC video encoder. The presented MPEG-4 AVC encoder was tested with test video sequences from the point of view of computational performance and coding efficiency. The runtime of the optimized video encoder is 37 to 132 times smaller relative to runtime of the reference MPEG-4 AVC encoder for comparable encoder compression performance.

We present the prototype of unmanned aerial vehicle (UAV) as a platform for multispectral acquisition. We are connecting the real-world simulation environment and control software to perform flight tests in SIL simulation. The full control system is based on the cascade of PI controllers with Anti-Windup mechanism, which stabilize the aircraft in the virtual reality. Stabilization of angular speed reduces problems connected with video disruptions. In this article we are presenting all implemented autonomous algorithms, which are based on ENU coordinate system (commonly used in aviation). Simulations are performed in Prepar3D® from Lockheed Martin which also allows to perform visual and thermal images processing. The prototype successfully completed all the test flights and is ready for various applications.

Currently, when Motion Capture being commonly used in gaming and movie industry there is a need of robust, easy and flexible solution to capture motion of non-humanoid characters in order to animate virtual characters in a game or movie. To fill this gap we developed an algorithm which estimates model of skeleton structure of both humanoid and non-humanoid characters. Quality and the possibility of real-life applications of the presented algorithm were experimentally evaluated. During the experiment we estimated the skeleton structure from the markers attached to a dog’s skin. Quality of the resulting model is very promising.

This paper presents a method for segmentation of 2D texture images of titanium alloys. The procedure is fully automated and is able to find and recognize so-called

α

-colonies from the image. The algorithm combines nonsubsampled directional filter banks (NSDFB) from the contourlet transform and gradient gray-level value to recognize directional orientations of

α

-colony.

Graph visualization is an ongoing research area with many open problems. Graphs are often visualized in 2D space, but recently also 3D visualizations emerged. However the added third dimension add additional problems. In this paper we focus on navigating and exploring 3D graph visualizations. We present our approach for the automation of virtual camera movement for better graph exploration that also allows to play-back the exploration and fork other exploration paths at any time.

The paper presents Improved Adaptive Arithmetic Coding algorithm for application in forthcoming HEVC video compression technology. The proposed solution is based on standard CABAC algorithm and uses author’s new mechanism of data statistics modeling that is based on CTW technique. The improved CABAC algorithm is characterized with better compression performance relative to standard CABAC. In the framework of HEVC encoder 1.6% - 4.5% bitrate reduction was obtained when using improved CABAC instead of the original algorithm.

This paper presents a method for extrinsic camera calibration (estimation of camera rotation and translation matrices) from a sequence of images. It is assumed camera intrinsic matrix and distortion coefficients are known and fixed during the entire sequence. Performance of the presented method is evaluated on a number of multi-view stereo test datasets. Presented algorithm can be used as a first stage in a dense stereo reconstruction system.

In this paper we discuss the problem of interpolating the so-called reduced data

$Q_m=\{q_i\}_{i=0}^m$

to estimate the length

d

(

γ

) of the unknown curve

γ

sampled in accordance with

γ

(

t

i

) = 

q

i

. The main issue for such non-parametric data fitting (given a fixed interpolation scheme) is to complement the unknown knots

$\{t_i\}_{i=0}^m$

with

$\{\hat t_i\}_{i=0}^m$

, so that the respective convergence prevails and yields possibly fast orders. We invoke here the so-called

exponential parameterizations (including centripetal)

combined with piecewise-quadratics (and -cubics). Such family of guessed knots

$\{\hat{t}_i^{\lambda}\}_{i=0}^m$

(with 0 ≤ 

λ

 ≤ 1) comprises well-known cases. Indeed, for

λ

 = 0

a blind uniform guess

is selected. When

λ

 = 1/2 the so-called

centripetal parameterization

is invoked. On the other hand, if

λ

 = 1

cumulative chords

are applied. The first case yields a bad length estimation (with possible divergence). In opposite, cumulative chords match the convergence orders established for the non-reduced data i.e. for

$(\{t_i\}_{i=0}^m, Q_m)$

. In this paper we show that, for exponential parameterization, while

λ

ranges from one to zero, diminishing convergence rates in length approximation occur. In addition, we discuss and verify a method of possible improvement for such decreased rates based on iterative knot adjustment.

In this paper we verify sharpness of the theoretical results concerning the asymptotic orders of trajectory approximation of the unknown parametric curve

γ

in arbitrary Euclidean space. The pertinent interpolation schemes (based on piecewise-quadratics and piecewise-cubics) are here considered for the so-called reduced data. The latter forms an ordered collection of points without provision of the associated interpolation knots. To complement such data i.e. to determine the missing knots, cumulative chord parameterization is invoked. Sharpness of cubic and quartic orders of convergence are demonstrated for piecewise-quadratics and piecewise-cubics, respectively. This topic has its ramification in computer vision (e.g. image segmentation), computer graphics (e.g. trajectory modeling) or in engineering (e.g. robotics).

Estimation of electrooculography (EOG) and blinking signals are important for medical applications and non–medical, like computer animation. Both signals are measured together using set of electrodes and the separation of them is necessary. Filter banks based technique for the estimation of EOG and blinking signals is considered. The gradient search allows estimation of the height and width of blinking pulses and the slopes between saccades. Improved detection related to the time domain of saccades is proposed too.

This paper presents a new way to use the well-known distortions introduced by the fisheye lens (ultra-wide-angle lenses) and a well-known algorithms for the correction of distortion in order to obtain completely new functionality in supervision and monitoring of video. The suggested way to use the barrel distortion will result in a smoother zooming and allows to achieve the multiple images simultaneously with the same resolution but with a different viewing angle. This feature may vary greatly increase the effectiveness of surveillance application by providing a greater amount of visual data from one camera.

In all hybrid video coders the raster scanning order of macroblock has been used. Such order scheme make parallel processing of macroblock very difficult. or even impossible The paper presents alternative method of macroblock ordering called “diamond scan”. This scheme of macroblock ordering allows for strong parallelization of macroblock processing and to have benefit from multicore processors (GP and GPU). Applying presented scheme leads to create a more efficient software and hardware implementations, in terms of energy consumption and efficient use of available equipment.

This paper describes self-quotient referring

ε

-filter for images under varying lighting conditions. Edge preserving smoothing is a fundamental feature extraction from the image for multimedia applications.

ε

-filter is a nonlinear filter, which can smooth the image while preserving edge information. The filter design is simple and it can effectively smooth the image. However, when we handle the image under light variation, the contrast of edge part is low in low contrast area, while it is high in high contrast area. Hence, the existing edge-preserving filters cannot preserve the edge information around low contrast area. Our method solves this problem by combining self-quotient filter and

ε

-filter. To confirm the effectiveness of the proposed method, we conducted some comparison experiments on face beautification.

A novel algorithm is proposed to locate the sets of adjacent co-spherical triangles for a given object, which enables us to detect spheres and spherical parts constituting the object. An extension of the idea of Hough transform has been used, aided by axial discretization and restricted searching, along with the geometric data structure of doubly connected edge list. The algorithm has been analyzed and shown to achieve significant efficiency in space and run-time. On testing the algorithm with various 3D objects, it is found to produce the desired result. Effects of different input parameters have been explained and the robustness of the algorithm has been shown for rough/noisy surfaces.

Here we are presenting a solution we developed for multimedia objects conversion in the digital repository that is being built under SYNAT/PASSIM project [1]. Our aim was to create tools for online format conversion of multimedia objects in a most simple way. We have studied available open source libraries and tools, and proposed an efficient way of integrating them into digital repository conversion module. Open structure of the formats supported by the constructed repository lead us to a plug-in based architecture. Here we are presenting the outline of our solution.

The wrist is the most common location for long-bone fractures in humans. To evaluate the healing process of such fractures, it is of interest to measure the fracture displacement, particularly the angle between the joint line and the long axis of the fractured long bone. We propose to measure this angle in 3D computed tomography (CT) images of fractured wrists. As a first step towards this goal, we here present a fast and precise semi-automatic method for determining the long axis of the radius bone in CT images. To facilitate user interaction in 3D, we utilize stereo graphics, head tracking, 3D input, and haptic feedback.

In the paper the Hybrid Feature Similarity metric is proposed based on the combination of two recently proposed objective image quality assessment methods - Riesz transform based Feature Similarity metric and Feature Similarity index. Both of them have good performance in comparison to most “state-of-the-art” quality metrics but highly linear correlation with subjective scores requires an additional nonlinear mapping for tuning to each dataset. In order to overcome this problem and obtain high quality prediction accuracy the nonlinear combination of both metrics is proposed leading to better performance than using each of the metrics separately. The experiments conducted in order to propose the weighting coefficients for both metrics have been performed using TID2008 dataset which is currently the largest and most comprehensive publicly available image quality assessment database, containing 1700 images together with their subjective quality evaluations. The verification of the obtained results has been also conducted using some other relevant benchmark databases.

Haptic devices are nowadays often used in non-hitech applications due to their increasing availability. These special input/output devices provide native 3D manipulation and, additionally, unlike mouse or keyboard, a control with a force. We present a low-level haptic visualization method which is able to simulate material on a surface and which works natively with the triangulated irregular network terrain models (TIN). Our method allows to model and visualize several materials, such as sand, concrete, wood, etc. New materials could be created by setting five parameters and, optionally, by the creation of a material relief. In our framework we use the TIN terrain model together with real-time erosion and interactive haptic editing using a set of virtual tools. We have performed a user study which is also included in the paper. Unlike existing methods, we use a general shape of the tool; our solution is capable of geometric editing of the TIN model, and it allows a simulation of several materials.

The paper presents some ways of distribution viewing points for generating viewing representation on viewing sphere with perspective, including methods using regular polyhedra and spiral path on sphere. The number of viewpoints follows from the required scanning resolution.

Image collections, both personal and commercial, are growing very rapidly. Consequently, methods for managing large image databases are highly sought after. In this paper, we look at various ways to visualise and interactively browse image collections. In general, we can divide image database visualisation approaches into three categories: mapping-based techniques which typically employ dimensionality reduction algorithms, clustered visualisations which group, often in a hierarchical manner, similar images, and graph-based approaches where links between images are exploited to arrive at an intuitive display of the dataset.

Once displayed, the user should be able to browse through the collection in an interactive, intuitive and efficient manner. Such browsing can be achieved in several ways. Horizontal browsing navigates through images of the same visualisation plane, and includes operations such as panning, zooming, magnification and scaling. In contrast, vertical browsing allows navigation to a different level of a hierarchically organised visualisation.

We present a fast, robust and fully automatic method for high dynamic range (HDR) images acquisition for non-static scenes. To obtain high correctness of the approach, perceptual experiments were conducted. HDR images became popular for realistic scene acquisition, as they register much more information than standard images. The most common approach for their acquisition is a composition of photographs taken with a conventional camera. However, the approach suffers from some limitations caused by even the smallest camera movements as well as by objects in motion in the scene. The last one causes ghost artifacts visible in a final image. The key components of our technique include probability maps calculated on the basis of sequences of hand-held photographs and perceptual experiments. We obtained validation of our results by HDR VDP technique.

This paper presents a new objective metric for assessing the visual difference between a reference or ‘perfect’ mesh and its distorted version. The proposed metric is based on the measurement of a distance between curvature tensors of the two triangle meshes under comparison. Unlike existing methods, our algorithm uses not only eigenvalues but also eigenvectors of the curvature tensor to derive a perceptually-oriented distance. Our metric also accounts for some important properties of the human visual system. Experimental results show good coherence between the proposed objective metric and subjective assessments.

This paper proposes a novel approach to recognize actions and objects within the context of each other. Assuming that the different actions involve different objects in image sequences and there is one-to-one relation between object and action type, we present a Bayesian network based framework which combines motion patterns and object usage information to recognize actions/objects. More specifically, our approach recognizes high-level actions and the related objects without any body-part segmentation, hand tracking, and temporal segmentation methods. Additionally, we present a novel motion representation, based on 3D Haar-like features, which can be formed by depth, color, or both images. Our approach is also appropriate for object and action recognition where the involved object is partially or fully occluded. Finally, experiments show that our approach improves the accuracy of both action and object recognition significantly.

This paper presents a new simple and robust set of features to classify emotional states in sequences of facial images. The proposed method is derived from simple geometric-based features that deliver a fast, highly discriminative, low-dimensional, and robust classification across individuals. The proposed method was compared to other state-of-the-art methods such as Gabor, LBP and AAM-based features. They were all compared using four different classifiers and experimental results based on these classifiers have shown that the proposed features are more stable in “leave-same-sequence-image-out” (LSSIO) environments, less computational intense and faster when compared to others.

Visual saliency models aim at predicting where people look. In free viewing conditions, people look at relevant objects that are in focus. Assuming blurred or out-of-focus objects do not belong to the region of interest, this paper proposes a significant improvement and the validation of a saliency model by taking blur into account. Blur identification is associated to a spatio-temporal saliency model. Bottom-up models are designed to mimic the low-level processing of the human visual system and can thus detect out-of-focus objects as salient. The blur identification allows decreasing saliency values on blurred areas while increasing values on sharp areas. In order to validate our new saliency model we conducted eye-tracking experiments to record ground truth of observer’s fixations on images and videos. Blur identification significantly improves fixation prediction for natural images and videos.

In this paper we present a method dedicated to hand tracking in sign language scenes using particle filtering. A a new penalisation method based on the optical flow mechanism is introduced. Generally, particle filters require the use of a reference model. In this paper we have introduced a new method based on a dictionary of visual references of hand to constitute the reference model. The evaluation of our method is performed on the SignStream-ASLLRP database on which we have provided ground truth annotations for this purpose. The obtained results show the accuracy of our method.

The paper presents a model to detect and track vehicles in highly congested traffic using low quality (usually compressed) video sequences. Robustness of the model is provided by applying a data fusion for various detection and tracking algorithms. The surveys to find reliable detection algorithms were performed. Basing on the experiments, the model calibration and results were presented. The proposed model provides data, which can be used by traffic engineers in various microscopic traffic simulations.

In this paper the approach to on-line object recognition for autonomous flying agent is considered. The method is divided into two parts. First the algorithm for scene objects vectorization is introduced. As the second step of the overall method we present the rotation and scale invariant algorithm for vectorized object identification based on syntactic language.

Geometrical limitations on the voting process in the classical Hough transform resulting from that the detected objects are opaque to the applied means of measurement are considered. It is assumed that the measurements are made from one point, like in LIDAR scanning. The detected object is a circle and the two point elementary voting set forming its chord is considered. The first type of conditions are those which can be used during the accumulation process. The

side condition

says that the circle lies at the opposite side of the chord than the laser source. The

magnitude condition

requires that points in the elementary set must not be occluded with respect to the source by any circle hypothesised in voting. The second type of conditions can be checked after after the detection. They require that points are neither inside the object not in its shadow. Departures from this rule are admitted, so fuzzy voting between positive and negative evidence for the object is considered.

Adaptive structuring elements modify their shape and size according to the image content and may outperform fixed structuring elements. Without any restrictions, they suffer from a high computational complexity, which is often higher than linear with respect to the number of pixels in the image. This paper introduces adaptive structuring elements that have predefined shape, but where the size is adjusted to the local image structures. The size of adaptive structuring elements is determined by the salience map that corresponds to the salience of the edges in the image, which can be computed in linear time. We illustrate the difference between the new adaptive structuring elements and morphological amoebas. As an example of its usefulness, we show how the new adaptive morphological operations can isolate the text in historical documents.

In our paper we discuss some of the problems of camera independent motion dection and propose the use of a qualitative method based on the projections of the optical flow. By applying several projections of the optical flow and using a voting mechanism we can increase the performance of motion detection: the F-measure, examined on 20 artificial and real-life test videos, was increased with about 10-25% in general compared to the average performance of individual projections.

The processing of the large amount of hand-written archive documents is an unsolved problem. We propose a semi-automatic text recognition approach for those documents containing a limited size of vocabulary. Our approach is word based and uses the Scale Invariant Feature Transform for finding and describing saliency points of hand-written words. For testing we used a book of a Central-European city census of the year 1771 containing mainly Christian and family names. At reasonable database size we could achieve about 80% recognition rate.

The paper presents a problem of recognition of color facial images in the aspect of dimensionality reduction performed by means of Principal Component Analysis employing different variants of input data organization. Here, input images are represented by tensors of 3rd order and the PCA is applied for matrices derived from such tensors. Its main advantage is associated with efficient representation of images leading to the accurate recognition. The paper describes practical aspects of the algorithm and its implementation for three variants of tensor unfolding. Furthermore the impact of the number of training/testing images, the reduction ratio and the color model on the recognition accuracy is investigated. The experiments performed on Essex facial image databases showed that face recognition using this type of feature space dimensionality reduction is particularly convenient and efficient, giving high recognition performance.

The paper presents a problem of recognition of facial portraits in the aspect of benchmark database quality. The aim of the work presented here was to analyse the potential of datasets published over the Internet and the predicted applicability of such data for the task of face recognition performance verification. We gathered 41 datasets created and published by various academic and commercial bodies. In the paper we focus on both pure data characteristics, including the number of images, their spatial resolution, quality, content and usability, as well as more high-level properties, e.g. face orientation, expression, background, lighting, and attributes like hats, glasses and beards. We have chosen several datasets on which we performed more detailed experiments related to face recognition. We employed several database preparation algorithms (cross-validation based on different schemes) to make the results as much objective as possible. Here, Principal Component Analysis was employed, as a standard tool for dimensionality reduction. The classification was performed using simple Euclidean metrics. Performed experiments showed a true potential of selected databases.

The use of the Polar-Fourier Greyscale Descirptor in the recognition of objects, which are very similar in shape is evaluated in the paper. For this purpose, a benchmark image database consisting of six butterflies species was applied. The investigated descriptor, which was designed for the representation of objects extracted from digital images, is based on the combination of polar and Fourier transforms applied for objects in greyscale. Some other operations are applied in order to improve the efficiency of the algorithm as a whole. The method was tested using 120 images of butterflies, 20 for 6 species, and has obtained 90% of efficiency.

The General Shape Analysis (GSA) is a problem of finding the most similar basic shape to the test one. It is close to traditional recognition or retrieval of shapes. Main difference is that GSA does not aim at the identification of an exact object shape but at the indication of one or few most similar to it general templates – simple shape figures, e.g. rectangle, circle or triangle. By comparing more complicated shapes with simple ones it is possible to determine the most general information about a particular object. In order to perform the comparison using the template matching approach it is necessary to define methods for the representation and similarity estimation of shapes. In this paper the attention is paid to two-dimensional Fourier Descriptor applied for the representation of a shape and two matching methods, namely Euclidean distance and correlation. The effectiveness of the shape descriptor is estimated as a convergence between the experimental results and results provided by humans through the inquiry forms concerning the same GSA task. Performed experiments allowed us to determine the influence of the matching method on the final effectiveness of the approach applying Fourier Descriptors. Selection of the absolute spectrum subpart size is also discussed.

Supervised pixel-based texture classification is usually performed in the feature space. We propose to perform this task in (dis)simil-arity space by introducing a new compression-based (dis)similarity measure. The proposed measure utilizes two dimensional MPEG-1 encoder, which takes into consideration the spatial locality and connectivity of pixels in the images. The proposed formulation has been carefully designed based on MPEG encoder functionality. To this end, by design, it solely uses P-frame coding to find the (dis)similarity among patches/images. We show that the proposed measure works properly on both small and large patch sizes. Experimental results show that the proposed approach significantly improves the performance of supervised pixel-based texture classification on Brodatz and outdoor images compared to other compression-based dissimilarity measures as well as approaches performed in feature space. It also improves the computation speed by about 40% compared to its rivals.

Road detection plays an important part in intelligent vehicle driving assistance system. In this paper, we present a real-time vision-based method which can detect drivable road area on unstructured urban roads. It first trains road models based on color cues from consecutive frames. Then, region growing method is employed on current frame to extract drivable areas with seeds points selected according to trained models. This method can adaptively detect drivable lane areas under normal and complicated road environment where there are shadows, lane markings or unstable lighting conditions. Experimental results on complex traffic scenes show that the proposed algorithm is effective and stable for real-time drivable road detection.

Semantically accurate segmentation of an object of interest (OOI) is a critical step in computer vision tasks. In order to bridge the gap between low-level visual features and high-level semantics, a more complete model of the OOI is needed. To this end, we revise the concept of directional spatial templates and introduce

regional

directional spatial templates as a means of including spatial relationships among OOI regions into the model. We present an object segmentation algorithm that learns a model which includes both visual and spatial information. Given a training set of images containing the OOI, each image is oversegmented into visually homogeneous regions. Next, Multiple Instance Learning identifies regions that are likely to be part of the OOI. For each pair of such regions and for each relationship, a regional template is formed. The computational cost of template generation is reduced by sampling the reference region with a pixel set that is descriptive of its shape. Experiments indicate that regional templates are an effective way of including spatial information into the model which in turn results in a very significant improvement in segmentation performance.

The paper proposes a method of controlling robotic manipulators with use of human gestures and movement. Experiments were performed with the use of 4 degree-of-freedom AX-12 Robotic Arm manipulator with force gripper and ASUS Xtion depth sensor also called motion controller. Depth and video capture has been done via OpenNI library. The infrastructure is based on Windows Communication Foundation (WCF) for remote access, authorization, multimedia streaming and servo control. Control of robotic manipulator is implemented with use of human computer interaction algorithm basing on depth sensor information.

A widely used pathological screening test for blood smears is the complete blood count which classifies and counts peripheral particles into their various types. We particularly interested in the classification and counting of the five main types of white blood cells (leukocytes) in a clinical setting where the quality of microscopic imagery may be poor. A critical first step in the medical analysis of cytological images of thin blood smears is the segmentation of individual cells. The quality of the segmentation has a great influence on the cell type identification, but for poor quality, noisy, and/or low resolution images, segmentation is correspondingly less reliable. In this paper, we compensate for less accurate segmentation by extracting features based on wavelets using the Dual-Tree Complex Wavelet Transform (DT-CWT) which is based on multi-resolution characteristics of the image. These features then form the basis of classification of white blood cells into their five primary types with a Support Vector Machine (SVM) that performs classification by constructing hyper-planes in a high multi-dimensional space that separates cases of different classes. This approach was validated with experiments conducted on poor quality, normal blood smear images.

Concealed weapon detection (CWD) is an important part of everyday law enforcement. There are numerous facilities that are endangered of an terrorist or an fanatic individual attack. Commercially used weapon detection gates are very expensive and sometimes impossible to install into already existing security infrastructures. Here we present a miniature prototype device for concealed weapon detection using two cameras: IR and visual. The prototype consists of two printed circuit boards (PCB). First PCB is responsible for analog to digital and digital to analog conversions of the video stream. The second board is the main processing unit realizing the presented fast image fusion algorithm. The relative size of the prototype can be assumed as a miniature in comparison to the current used solutions. Such miniature device could be mounted under the ceiling or inside 3 DOF gimbals for wider view angle. Presented device can be considered as an alternative to already existing man-sized gates traditionally used for CWD.

Studying changes in the cell after treatment by a potentially genotoxic agent is a very important approach in biological experimental techniques. A researcher in molecular biology can study the structure of cellular pathways related to responses to external agents stress, damage or ionizing radiation by measuring the amount of cellular species, such as certain proteins, RNA or DNA. For this aim the electrophoresis-based tests are widely used. In this paper we will focus on the Western blot assay for estimating the quantity of the proteins. Often such information is obtained by visual analysis of a gel, which is prejudicial and time-consuming. We developed a new, rapid and exact image processing method for automatic detection of the spots in 2-D gels and calculation the quantity of the protein produced by the cell. The proposed method can significantly reduce the time required for analysis. We have obtained very promising results with accuracy more than 96% allowing for automated analyzes of 2-D gel electrophoresis images.

A model-based object recognition in video and depth images is proposed for the purpose of semantic map creation in mobile robotics. Three types of objects are modeled: a human silhouette, a chair/table and corridor walls. A bi-driven hypothesis generation and verification strategy is outlined. The object model includes a hierarchic semantic nets, combined with a graph of constraints and a Bayesian network for hypothesis generation and evaluation. For the purpose of model-to-image matching we define an incomplete constraint satisfaction problem and solve it. Our CSP-search allows partial assignment solutions and uses a stochastic inference to provide judgments of such solutions. The verification of hypotheses is due to a top-down occlusion propagation process, that explains why some object parts are hidden or occluded.

Skin regions detection has been intensively studied and many methods were proposed which are based on skin color modeling in different color spaces. This makes it possible to transform color images into skin probability maps and extract skin regions. However, in very few cases spatial alignment of the skin pixels is taken into account. In this paper we present how the pixel-wise detectors can be improved using distance transform performed in a combined domain of the skin probability maps and luminance. The proposed method is compared theoretically and experimentally with a well-established controlled diffusion technique for determining skin regions from skin probability maps.

Depth is very useful cue to achieve reliable fall detection since humans may not have consistent color and texture but must occupy an integrated region in space. In this work we demonstrate how to accomplish reliable fall detection using depth image sequences. The depth images are extracted by low-cost Kinect device. The person undergoing monitoring is extracted through mean-shift clustering. A depth connected component algorithm is used to delineate he/she in sequence of images. The system permits unobtrusive fall detection as well as preserves privacy of the user. The experimental results indicate high effectiveness of fall detection in indoor environments and low computational overhead of the algorithm.

Dimensionality reduction is a big challenge in many areas. In this research we address the problem of high-dimensional hyperspectral images in which we are aiming to preserve its information quality. This paper introduces a study stability of the non parametric and unsupervised methods of projection and of bands selection used in dimensionality reduction of different noise levels determined with different numbers of data points. The quality criteria based on the norm and correlation are employed obtaining a good preservation of these artificial data in the reduced dimensions. The added value of these criteria can be illustrated in the evaluation of the reduction’s performance, when considering the stability of two categories of bands selection methods and projection methods. The performances of the method are verified on artificial data sets for validation. An hybridization for a better stability is proposed in this paper, Band Clustering (BandClust) with Multidimensional Scaling (MDS) for dimensionality reduction. Examples are given to demonstrate the hybridization originality and relevance(BandClust/MDS) of the analysis carried out in this paper.

This paper proposes to revisit a recent interactive segmentation algorithm based on an original image representation called the component-tree [1]. This method relies on an optimisation process allowing to choose a segmentation result fitting at best some image markers defined by the user. We propose different solutions to improve the efficiency of the method, in particular by including meaningful photometric informations and by assessing automatically the user parameter

α

.

During breast cancer malignancy grading the main problem that has direct influence on the classification is imbalanced number of cases of the malignancy classes. This poses a challenge for pattern recognition algorithms and leads to a significant decrease of the classification accuracy for the minority class. In this paper we present an approach which ameliorates such a problem. We describe and compare several state of the art methods, that are based on the oversampling approach, i.e. introduction of artificial objects into the dataset to eliminate the disproportion among classes. We also describe the automatic thresholding and fuzzy c-means algorithms used for the nuclei segmentation from fine needle aspirates. Based on the segmented images a set of 15 feattures used for classification process was extracted.

We present an algorithm for view-independent human gait recognition. The human gait recognition is achieved using data obtained by our markerless 3D motion tracking algorithm. The tensorial gait data were reduced by multilinear principal component analysis and subsequently classified. The performance of the motion tracking algorithm was evaluated using ground-truth data from MoCap. The classification accuracy was determined using video sequences with walking performers. Experiments on multiview video sequences show the promising effectiveness of the proposed algorithm.

This paper introduces a new approach for gait-based gender classification in which some key biomechanical poses of a gait pattern are represented by partial Gait Energy Images (GEIs). These pose-based GEIs can more accurately represent the shape of the body parts and some dynamic features with respect to the usually blurred depiction provided by a general GEI comprising all poses. Gait-based gender classification is based on the weighted decision fusion of the pose-based GEIs. Results of experiments on two large gait databases prove that this method performs significantly better than clasiffiers based on the original GEI.

The paper presents a comparison of key point selection methods used for recognition of objects in scenes recorded by a built-in mobile phone camera. The detected key points include corners and line crossings. An application for Android smartphones was developed utilizing the Features from Accelerated Segment Test (FAST) and Scale-Invariant Feature Transform (SIFT), which was specially modified for processing low resolution images. The implemented algorithm computes descriptors which are invariant to image acquisition settings such as: rotation, noise, scale and brightness variations. The proposed image classification algorithm is based on pairing key points based on similarity of their descriptors.

Image Based Lighting technique needs light probe images. Light probe images are measurements of the scene light. Spherical and hemispherical mirrors (light probe measurement devices) and camera are used for the light probe image acquisition. In the paper is proposed and analyzed computational requirement of position and radius estimation of the hemispherical mirror with stripe pattern flange. Proposed solution reduces computation cost and allows processing of 4k image in 3 minutes.

The scope of this article is gait identification of individuals on the basis of reduced sequences of video recordings data. The gait sequences are considered to be the 3rd-order tensors and its dimensionality is reduced by Multilinear Principal Component Analysis with different values of variation covered. Reduced gait descriptors are identified by the supervised classifiers: Naive Bayes and Nearest Neighbor. CASIA Gait Database ’dataset A’ is chosen to verify the proposed method. The obtained results are promising. For the Naive Bayes and attributes discretization almost 99% of classification accuracy is achieved, which means only one misclassified gait out of eighty validated.

In this paper the novel modification of the well known Canny edge detection algorithm is presented. The first section describes the goal to be achieved by using the new algorithm. The second section describes theoretical basis of Canny algorithm and its practical implementation. Next, basics of the Ramer–Douglas–Peucker algorithm used for reducing the number of points in the curve are presented. The extension of the Canny algorithm and its implementation are presented in the fourth section. The next section shows the results of the new algorithm implementation for various images and presents statistical data to report effectiveness of the proposed algorithm modification.

This paper addresses the problem of providing traffic data for traffic control systems especially local traffic controllers, which optimize control sequences based on traffic loads at intersections. Optimization procedures require reliable data on preceding traffic changes for calculation of control commands. 3D wavelet representation of the road image sequence is proposed for use as an equivalent of traffic stream. Coefficients of this representation map with sufficient accuracy such traffic parameters as traffic density, traffic flow intensity and derivates. The level of wavelet decomposition is determined by the size and speed of the observed objects. Computation of the wavelet transform (3D DWT) may be easily performed using logic based circuits, which is an attractive solution for incorporation into local traffic controllers.

In the present paper an algorithm for the detection of voids of dental root canal obturation in microCT images of the root is described. The algorithm consists of segmentation of the filling material, based on histogram analysis, detection of the voids surrounded by the filling material, detection of voids at the interface of dentine and root canal filling and final processing. The segmentation requires selecting two threshold levels and involves histogram thresholding followed by region growing. To detect the voids at the interface of the filling material a variant of a hit-or-miss filter is proposed. The performance of the algorithm is tested, based on a set of microCT images.

This paper demonstrates the application of the Swarm Intelligence (SI) algorithm to recognize the specific patterns that are present in the digital images of handwritten music scores. The application introduced in this paper involves the detection of stafflines using particle swarm. The introduced solution described in this paper is a new approach to the problem, and illustrates how optimization algorithm can be modified and successfully applied in different subjects such as pattern recognition. The developed algorithm can be used as a first stage in Optical Music Recognition (OMR) that is followed by the staffline removal phase. It is worth pointing out, that contrary to most state-of-the-art algorithms, the proposed method does not require a binarization step in the preprocessing stage.

The paper presents an algorithm for building a map of obstacles and guiding an autonomous mobile platform in an unknown and changing environment. Depth images captured from a stereovision camera are used to detect objects and denote their location on the obstacle map. The depth images acquired from the stereocamera are encumbered with artefacts which poses the main problem in detecting obstacles. We propose a two-step filtering algorithm which is based on morphological operations and Bayesian inference. Experimental results proved the efficiency of the solution in the real environment wherein both static and mobile obstacles are present.

Development of human-computer interaction methods tends to exploit more and more natural human activities like thoughts, body posture or hands gesticulation. While most of authors improve whole body tracking this paper concentrates on hand’s poses analysis. Due to the usage of the depth image based object recognition approach to hand pose estimation a very precise method was obtained. Additionally thanks to decision forest implemented on GPU a real-time processing is possible.

Robust lip contour detection plays an important role in Facial Expression Recognition (FER). However, the large variations emerged from different speakers, intensity conditions, poor texture of lips, weak contrast between lip and skin, high deformability of lip, beard, moustache, wrinkle, etc. often hamper the lip contour detection accuracy. The novelty of this research effort is that we propose a new lip boundary localization scheme using Game Theory (GT) to elicit lip contour accurately from a facial image. Furthermore, we apply a feature subset selection scheme based on Particle Swarm Optimization (PSO) to select the optimal facial features. We have conducted several sets of experiments to evaluate the proposed approach. The results show that the proposed approach has achieved recognition rates of 93.0% and 92.3% on the JAFFE and CK+ datasets, respectively.

This paper presents a multibiometric system that integrates the iris, palmprint, and fingerprint features based on the fusion at feature level. The novelty of this research effort is that we propose a feature subset selection scheme based on Particle Swarm Optimization (PSO) with a new fitness function that minimizes the Recognition Error (RR), False Accept Rate (FAR), and Feature Subset Size (FSS). Furthermore, we apply a Distance Regularized Level Set (DRLS)-based iris segmentation procedure, which maintains the regularity of the level set function intrinsically during the curve evolution process and increases the numerical accuracy substantially. The proposed iris localization scheme is robust against poor localization and weak iris/sclera boundaries. Experimental results indicate that the proposed approach increases biometric recognition accuracies compared to that produced by single modal biometrics.

This paper presents a new near neighbor search. Feature vectors to be stored do not have to be of equal length. Two feature vectors are getting compared with respect to supremum norm. Time demand to learn a new feature vector does not depend on the number of vectors already learned. A query is formulated not as a single feature vector but as a set of features which overcomes the problem of possible permutation of components in a representation vector. Components of a learned feature vector can be cut out - the algorithm is still capable to recognize the remaining part.

A novel method of prostate segmentation in a new CT data making use of explicit knowledge about the prostate is proposed. The segmentation procedure is based on active shape statistical model (ASM) of the prostate, calculated using available data base of CTs annotated by medical doctors. In the paper the problem of automatic calculation of

corresponding

prostate landmarks in

different

CTs, which are absolutely necessary for the ASM, is solved in a new manner by: 1) finding parameters of affine and B-spline transformations in groupwise registration framework ensuring pixel-based registration of all available CTs in one common co-ordinate system, 2) performing forward affine and B-spline transformation of the annotated prostate contours into this co-ordinate system, 3) averaging them - interpolation & re-sampling, 4) propagation (projection) of mean landmarks, obtained in common co-ordinate system, to the training CTs using the backward transformation. Having the same prostate landmarks in set of CTs, the ASM of the prostate is calculated (its mean shape and tendencies to its direction variations). The result of matching ASM to the data is treated as the prostate segmentation result. Obtained results are presented and discussed in the paper.

Resting tremor is one of the primary motor symptoms of Parkinson’s Disease. In this paper we analyze the occurrence of tremor in Parkinson’s disease by using a system for measurements of kinematic data of upper limbs. The experimental group includes seven PD patients during standing. Analysis is based on kinematic measurements done by using multimodal motion capture (MOCAP) system for registration of 3D positions of body markers, ground reaction forces and electromyography signals. All patients taking part in examination have undergone deep brain stimulation surgical treatment. Examination involved comparisons of tremor parameters across four conditions, where stimulator was turned ON/OFF and medication was ON/OFF. Obtained results confirm statistically significant differences of certain tremor parameters between different experimental conditions.

An integrated segmentation approach for color images and depth maps is proposed. The 3D pointclouds are characterized by normal vectors and then grouped into planar, concave or convex faces. The empty regions in the depth map are filled by segments of the associated color image. In the experimental part two types of depth maps are analysed: generated by the MS-Kinect sensor or by a stereo-pair of cameras.

The detection of comic panels is a crucial funcionality in assistance systems for iconotextual media analysis. Most systems use recursive cuts based on image projections or background segmentation to find comic panels. Usually this limits the applicability to comics with white background and free space between the panels. In this paper, we introduce a set of new features that allow for a detection of panels by their outline instead of the separating space. Our method is therefore more tolerant against structured backgrounds.

The aim of this research is to develop a method for classification of the degree of myocardial necrosis using texture parameters estimated for static ultrasound images. The study is performed for the color and monochrome contrast echocardiograms that allow the advanced evaluation of myocardial function. The analysis includes investigation of different texture feature selection methods and application of two neural networks with different architectures along with SVM for classification. The obtained preliminary results are promising; classification error in all investigated cases is lower than 20%. The results were presented and discussed, also direction of further research was outlined.

In this paper the analysis of the premature born babies abdominal blood oxidation values as a signal is conducted with the goal of acquisition of the basic parameters of the signal and establishing the reference parameters for further studies. The authors also study the behavior of the signal and determine the possibility of its prediction using ARIMA model. To authors’ knowledge no such analysis of the signal from preterm babies was conducted yet, both from medical and computer science points of view, so in this paper they also try to answer the question whether the signal may be reliable for further studies on the possible use of it in monitoring and diagnosis of the preterm babies.

The representation and the thorough understanding of human motion is a crucial and challenging problem which has been raised in many scientific areas. This paper considers approaches in performing motion analysis with multi-resolution techniques based on rotations of joints over the time written in the form of a quaternion signal. The second generation wavelet transform constructed by the lifting scheme for the quaternion rotation representation can be used. Quaternions in terms of motion analysis are a more efficient representation of rotation than Euler angles. This paper presents the new quaternion lifting scheme building blocks for the smooth second degree transform based on the spherical cubic quaternion interpolation method (SQUAD). Also the possible applications of result multi-resolution representation as feature detection and compression are described.

The goal of this paper was to verify whether the conditional random fields are suitable and enough efficient for eye blink detection in user authentication systems based on face recognition with a standard web camera. To evaluate this approach several experiments were carried on using a specially developed test application and video database.

Morphological feature extraction allows obtaining a feature vector that can be used in pattern recognition. It is a two stage process, based on the extraction of morphological spatial classes and class distribution functions in order to obtain a feature vector. In this paper, we discuss two ways of considering directionality within this process. The first approach is based on division of the image space into sectors, in which the spatial classes are computed. The second makes use of directional structuring element used by morphological operators. Example applications and test results are also presented in the paper.

Modern livestock farming follows a trend to higher automation and monitoring standards. Novel systems for the health monitoring of animals like dairy cows are under development. The application of infrared thermography (IRT) for medical diagnostics was suggested long ago, but the lack of suitable technical solutions still prevents an efficient use. Within the R&D project VIONA new solutions were developed to provide veterinary IRT based diagnostic procedures. Therefore a reliable object detection and segmentation of the IR images is required. Due to the significant shape variation of the objects of interest advanced segmentation methods are necessary. The level set approach is applied to veterinary IR images for the first time. The special features of the thermal infrared spectrum require extensive adaptations of the approach. The suggested probability based shape prior and results of the successful application on IR images of dairy cows are presented.

Clustering in 2D space can be adapted as a segmentation method in images. In this study, we improve one well-known clustering algorithm, DBSCAN, to tackle pattern recognition problems in natural images. In DBSCAN, the details of objects are lost because of the noise in the scene or boundary regions. We overcome this problem using multi-scale approach to collect the salient features at different scales for better clustering. We use Gaussian kernel to smooth an image since multi-scale approaches are shown to be a well modeled with this kernel. Comparing with manually segmented images as gold standard, we show that the proposed multi-scale framework outperforms the segmentation of objects obtained with DBSCAN.

In this paper, we propose a tree-based approach to represent and compare image objects. Upon objects separated from images trees are constructed. The key observation is that from similar objects similar trees are produced. On the other hand, upon dissimilar objects unlike trees are created. Additionally, the degree of dissimilarity between objects is proportional to the degree of dissimilarity between the trees. Hence, it is possible to express the difference between two objects as the difference between the trees. The paper presents algorithms of creating and comparing trees as well as results, which confirm usefulness of the approach.

This work introduces a 3D+t coronary registration strategy to minimize the navigation among cardiac phases during the process of ischaemic heart disease diagnosis. We propose to register image sub-volumes containing suspected arterial lesions at two cardiac phases, instead of performing a registration of the complete cardiac volume through the whole cardiac cycle. The method first automatically defines the extent of the sub-volumes to be aligned, then the registration is performed in two steps: a coarse rigid alignment and a deformable registration. Our method provides comparable results and is computationally less expensive than previous approaches that make use of larger spatial and temporal information.

We observe increasing complexity of Information Systems that is seemingly in contrast to constant limits of human perception. In order to handle this intricacy, various automated or semi-automated methods for data visualization are implemented. Graphical User Interface design is very important for human-computer interaction since it improves productivity and enhances human understanding. This paper proposes the novel algorithm for estimation and deployment of the visual elements of the GUI layout. The layout is composed of the blocks, that eventually are replaced by the system dependent object representations, e.g. visual metaphors of the heating system. The paper discusses the theoretical background, the properties of proposed algorithms together with the sample prototype application.

The paper provides fundamental information about the SmartMonitor – an innovative surveillance system based on video content analysis. We present a short introduction to the characteristics of the developed system and a brief review of methods commonly applied in surveillance systems nowadays. The main goal of the paper is to describe planned basic system parameters as well as to explain the reason for creating it. SmartMonitor is being currently developed but some experiments have already been performed and their results are provided as well.

This paper presents a new image denoising method based on contourlet transform and Lee filter. Classical contourlet transform methods are based on denoising procedure that processes the contourlet coefficients with a threshold in each subband. This is performed without considering the neighbourhood characteristics of the invariance of the contourlet transform which introduces some artifacts. In this work, however, Lee filter is used to solve this problem. The suggested algorithm is particularly useful when considering biometric images that need precise preprocessing.

In this paper, we present an approach for tackling the problem of automatically detecting and tracking a varying number of people in complex scenes. We follow a robust and fast framework to handle unreliable detections from each camera by extensively making use of multi-camera systems to handle occlusions and ambiguities. Instead of using the detections of each frame directly for tracking, we associate and combine the detections to form so called tracklets. From the triangulation relationship between two views, the 3D trajectory is estimated and back-projected to provide valuable cues for particle filter tracking. Most importantly, a novel motion model considering different velocity cues is proposed for particle filter tracking. Experiments are done on the challenging dataset PETS’09 to show the benefits of our approach and the integrated multi-camera extensions.