Bogusław Cyganek

Clustering-based ensembles for one-class classification

This paper presents a novel multi-class classifier based on weighted one-class support vector machines (OCSVM) operating in the clustered feature space. We show that splitting the target class into atomic subsets and using these as input for one-class classifiers leads to an efficient and stable recognition algorithm. The proposed system extends our previous works on combining OCSVM classifiers to solve both one-class and multi-class classification tasks. The main contribution of this work is the novel architecture for class decomposition and combination of classifier outputs. Based on the results of a large number of computational experiments we show that the proposed method outperforms both the OCSVM for a single class, as well as the multi-class SVM for multi-class classification problems. Other advantages are the highly parallel structure of the proposed solution, which facilitates parallel training and execution stages, and the relatively small number of control parameters.

One-Class Support Vector Ensembles for Image Segmentation and Classification

This paper presents an extension of the one-class support vector machines (OC-SVM) into an ensemble of soft OC-SVM classifiers. The idea consists in prior clustering of the input data with a kernel version of the deterministically annealed fuzzy c-means. This way partitioned data is trained with a number of soft OC-SVM classifiers which allow weight assignment to each of the training data. Weights are obtained from the cluster membership values, computed in the kernel fuzzy c-means. The method was designed and tested mostly in the tasks of image classification and segmentation, although it can be used for other one-class problems.

Hybrid computer vision system for drivers' eye recognition and fatigue monitoring

This paper presents a hybrid visual system for monitoring drivers states of fatigue, sleepiness and inattention based on drivers eye recognition. Safe operation in car conditions and processing in daily and night conditions are obtained thanks to the custom setup of two cameras operating in the visible and near infra-red spectra, respectively. In each of these spectra image processing is performed by a cascade of two classifiers. The first classifier in a cascade is responsible for detection of eye regions based on the proposed eye models specific to each spectrum. The second classifier in each cascade is responsible for eye verification. It is based on the higher order singular value decomposition of the tensors of geometrically deformed versions of real eye prototypes, specific to the visible and NIR spectra. Experiments were performed in real car conditions in which four volunteer drivers participated. The obtained results show high recognition accuracy and real-time processing in software implementation. Thanks to these the system can become a part of the advanced drivers assisting system.

Comparison of nonparametric transformations and bit vector matching for stereo correlation

The paper describes and compares stereo matching methods based on nonparametric image transformations. The new nonparametric measures for local neighborhoods of pixels are proposed as well. These are extensions to the well known Census transformation, successively used in many computer vision tasks. The resulting bit-fields are matched with the binary vectors comparison measures: Hamming, Tanimoto and Dixon-Koehler. The presented algorithms require only integer arithmetic what makes them very useful for real-time applications and hardware implementations. Many experiments with the presented techniques, employed to the stereovision, showed their robustness and competing execution times.

Image recognition with deep neural networks in presence of noise – Dealing with and taking advantage of distortions

Data classification in presence of noise can lead to much worse results than expected for pure patterns. In this paper we investigate this problem in the case of deep convolutional neural networks in order to propose solutions that can mitigate influence of noise. The main contributions presented in this paper are experimental examination of influence of different types of noise on the convolutional neural network, proposition of a deep neural network operating as a denoiser, investigation of a deep network training with noise contaminated patterns, and finally an analysis of noise addition during the training process of a deep network as a form of regularization. Our main findings are construction of the deep network based denoising filter which outperforms state-of-the-art solutions, as well as proposition of a practical method of deep neural network training with noisy patterns for improvement against the noisy test patterns. All results are underpinned by experiments which show high efficacy and possibly broad applications of the proposed solutions.

Road Signs Recognition by the Scale-Space Template Matching in the Log-Polar Domain

This paper presents a cascaded system for recognition of the circular road-signs. The system consists of two compound detectors-classifiers. Each operates on the Gaussian scale-space and does template matching in the log-polar domain. The first module is responsible for detection of the potential sign areas at the coarsest level of the pyramid. The second one, in turn, refines the already found places at the finest level. Thanks to this composition, as well as to the efficient matching in the log-polar domain, the system is very robust in terms of recognition of the signs with different scales and rotations, as well as under partial occlusions, poor illumination conditions, and noise.

Real-time detection of the triangular and rectangular shape road signs

Road signs recognition systems are developed to assist drivers and to help increase traffic safety. Shape detectors constitute a front-end in majority of such systems. In this paper we propose a method for robust detection of triangular, rectangular and rhombus shaped road signs in real traffic scenes. It starts with segmentation of colour images. For this purpose the histograms were created from hundreds of real warning and information signs. Then the characteristic points are detected by means of the developed symmetrical detector of local binary features. The points are further clusterized and used to select shapes from the input images. Finally, the shapes are verified to fulfil geometrical properties defined for the road signs. The proposed detector shows high accuracy and very fast operation time what was verified experimentally.

An improved variogram analysis of the maximum expected disparity in stereo images

This paper concerns the method of estimation of one of the initial parameters for stereo image processing - the maximum expected disparity value. An automatic assessment of this parameter would benefit in improvement of automation and performance of stereo methods. The authors have developed an improved version of the heuristic method of estimation of the maximum disparity value for real stereo images. It is based on statistical analysis of the spatial correlation between stereo images - the so called image variograms, used so far only to single images and extended to the processing of stereo views by authors of this paper.

Recognition of road signs with mixture of neural networks and arbitration modules

The automatic detection and recognition of road signs play important role in the driver assistance systems and can increase the safety on the roads. In this paper we propose a system of a road signs classifier which is based on ensemble of the non Euclidean distance neural networks and an arbitration unit. The input to this system comes from the sign detection module which supplies a normalized, binarized and resampled pictogram of a detected sign. The system performs classification on deformable models. The classifier is composed of a mixture of experts (binary distance neural networks) operating on slightly tilted or shifted versions of pictograms. This ensemble of experts is orchestrated by an arbitration module which operates in the winner-takes-all mode with a novel modification of promoting the most populated group of unanimous experts. The experimental results showed great robustness of the system and very fast response time which is an important factor in the driving assistance systems.

Multidimensional data classification with chordal distance based kernel and Support Vector Machines

In contemporary machine learning multidimensional rather than pure vector like data are frequently encountered. Traditionally, such multidimensional objects, such as color images or video sequences, are first transformed to a vector representation, and then processed by the classical learning algorithms operating with vectors. However, such multi-to-one dimension transformations usually lead to loss of important information. Thus, proposing novel methods for representing and learning with complex and multidimensional data is in focus of current machine learning research. In this paper, we propose a new method for efficient classification of multidimensional data based on a tensor-based kernel applied to the Support Vector Machines. We represent data as tensors, in order to preserve data dimensionality and to allow for processing of complex structures. To allow for an effective classification, we augment a Support Vector Machine (SVM) trained with Sequential Minimal Optimization (SMO) procedure with a chordal distance-based kernel for efficient classification of tensor-like objects. We also discuss different optimization methods for SVM, as well as present implementation details with computational time analysis. The proposed method is evaluated in both binary and multi-class classification problems. Comprehensive experimental analysis carried on a number of multidimensional benchmarks shows high usefulness of the proposed approach.