Mihaela Gordan

A support vector machine-based dynamic network for visual speech recognition applications

Visual speech recognition is an emerging research field. In this paper, we examine the suitability of support vector machines for visual speech recognition. Each word is modeled as a temporal sequence of visemes corresponding to the different phones realized. One support vector machine is trained to recognize each viseme and its output is converted to a posterior probability through a sigmoidal mapping. To model the temporal character of speech, the support vector machines are integrated as nodes into a Viterbi lattice. We test the performance of the proposed approach on a small visual speech recognition task, namely the recognition of the first four digits in English. The word recognition rate obtained is at the level of the previous best reported rates.

Application of support vector machines classifiers to visual speech recognition

In this paper we propose a visual speech recognition network based on support vector machines. Each word of the dictionary is modeled by a set of temporal sequences of visemes. Each viseme is described by a support vector machine, and the temporal character of speech is modeled by integrating the support vector machines as nodes into a Viterbi decoding lattice. Experiments conducted on a small visual speech recognition task using very simple features demonstrate a word recognition rate on the level of the best rates previously reported even without training the state transition probabilities in the Viterbi lattices. This proves the suitability of support vector machines for visual speech recognition.

Structured light self-calibration with vanishing points

This paper introduces the vanishing points to self-calibrate a structured light system. The vanishing points permit to automatically remove the projector’s keystone effect and then to self-calibrate the projector–camera system. The calibration object is a simple planar surface such as a white paper. Complex patterns and 3D calibrated objects are not required any more. The technique is compared to classic calibration and validated with experimental results.

A fuzzy optimization method for CMOS operational amplifier design

The aim of the paper is to present a fuzzy method for the optimization of the CMOS operational amplifier design. Our method uses fuzzy systems or fuzzy sets in all stages involved in the optimization process. In order to reduce the time spent for circuit performance evaluation, we use fuzzy system to model each circuit performance. The optimization problem formulation is accomplished in a flexible manner using fuzzy sets to define fuzzy optimization objectives. We use qualitative design knowledge to modify the design parameters in each iteration. This is done using a fuzzy system for each parameter. After introducing our fuzzy optimization method we design a basic two-stage CMOS operational amplifier.

3D DCT supervised segmentation applied on liver volumes

Liver segmentation from computer tomography scans is a topic of research interest, as the acquisition and inter-patient variability make the automatic segmentation difficult. The current trend is to improve the accuracy and to reduce the computational complexity of the segmentation, as this is essential for the diagnostic and for 3D rendering. We propose a new computationally efficient approach for 3D liver segmentation, based on the 3D Discrete Cosine Transform applied on volume blocks for feature extraction, followed by a support vector machine classification of volume blocks. The segmentation is refined in a post-processing step through a 3D median filtering, 3D morphological operations, and 3D connected components analysis. This new method has been applied on real liver volumes and provided promising results, on the level of the state of the art, with a significant reduction in the data to be processed and in the operations involved as compared to other approaches.

Retinal vessels segmentation using supervised classifiers decisions fusion

Ophthalmology is a significant branch of the biomedical field which requires computer-aided automated techniques for pathology identification. Within this framework, an important concern is the accurate segmentation of the retinal blood vessels. A reference approach in the literature to this task consists in the classification of the pixels as vessels or non-vessels, using as discriminative features the green channel intensity, two-dimensional Gabor wavelet responses and some variants of LBP descriptors. However the discriminative power of this feature set is not always sufficient to provide a really highly accurate segmentation. In this paper we propose a new approach, combining powerful machine learning classifiers: support vector machines and neural networks over the same feature set, to improve the classification accuracy by a weighted decision fusion. The experimental results obtained on the DRIVE database show that the segmentation accuracy is increased up to 94%, which is superior to similar segmentation methods from the literature using neural networks, Bayesian, unsupervised classifiers and even support vector machines individually. When these results are further combined with the output of matched filters applied on the retinal images, the segmentation accuracy is further increased, by a better identification of the fine vessels.

Fuzzy intensification operator based contrast enhancement in the compressed domain

With the increasing sizes of high resolution images, their storage and processing directly in the compressed domain has significantly gained importance. Algorithms for compressed domain image processing provide a powerful computational alternative to classical (pixel level) based implementations. While linear algorithms can be applied straightforward to the JPEG compressed images, this is not the case for nonlinear image processing, as for example contrast enhancement algorithms. In this paper a new implementation in the compressed domain of a very efficient contrast enhancement, based on fuzzy set modeling and on a fuzzy intensification operator, is presented. The fuzzy set parameters are adaptively chosen by analyzing the statistics of the image data in the compressed domain, in order to optimally enhance the image contrast. The nonlinear enhancement procedure requires a grey level threshold, for which an adaptive implementation, taking into account the frequency content of each coefficient block in the DCT (Discrete Cosine Transform) encoded JPEG image is proposed. This guarantees the optimal quality at minimum computational cost. The experimental results for a set of various contrast images validate the good performance and functionality of the proposed implementation.

A New SVM-Based Architecture for Object Recognition in Color Underwater Images with Classification Refinement by Shape Descriptors

Underwater images analysis is a difficult task due to their specific attributes: weak and variable lighting, low contrast, blurring. Therefore powerful image analysis algorithms, application specific, must be employed to obtain good results. In this paper we propose such a novel architecture based on a support vector machine (SVM) classifier, dedicated to large underwater scenes analysis for the specific task of localizing circular shaped objects of known dimension - the pressure equalization openings on the underwater face of a hydro-dam. Despite the very good recognition performance of SVM classifiers, their success on such poor quality images is relatively modest. The new proposed architecture achieves the maximization of the correct classification rate in two steps. In the first step, a non-linear SVM classifier is trained on raw color pixel features extracted from regions of interest of approximately the objects size. In the classification phase, the underwater image to be analyzed is decomposed on such partially overlapping elementary regions of interest. The regions of interest are classified by the SVM, using as threshold for the decision function a real value selected to minimize the false rejection rate. Then, in the second step, a shape descriptor (the circularity) of the patterns classified by the SVM as objects of interest is computed. This shape descriptor of the positive patterns is used for their classification as objects of interest or not, through a simple threshold comparison. As a result of this classification, the false acceptance rate is minimized as well, thus refining the classification results

Visual speech recognition using support vector machines

In this paper we propose a visual speech recognition network based on support vector machines. Each word of the dictionary is described as a temporal sequence of visemes. Each viseme is described by a support vector machine, and the temporal character of speech is modeled by integrating the support vector machines as nodes into a Viterbi decoding lattice. Experiments conducted on a small visual speech recognition task show a word recognition rate on the level of the best rates previously reported, even without training the state transition probabilities in the Viterbi lattice and using very simple features. This proves the suitability of support vector machines for visual speech recognition.

Pseudoautomatic lip contour detection based on edge direction patterns

Detection and tracking of the lip contour is an important issue in lipreading. While there are solutions for lip tracking once a good contour initialization in the first frame is available, the problem of finding such a good initialization is not yet solved automatically, but done manually. Solutions based on edge detection and tracking have failed when applied to real world mouth images. In this paper, we propose a solution to lip contour detection that minimizes user interaction by requiring a minimal number of points to be marked manually on the mouth image. The proposed approach is based on edge detection using gradient masks and edge following. The method is based on the examination of gradient direction patterns in the lip area, and makes use of the local direction constancy along the lip contours, as opposed to the other regions of the mouth image that are characterized by random edge directions.