Earticle

This paper presents a new Speaker Recognition Technique aimed at high identification accuracy and low impostor acceptance. This method is based on a modified neural network, which is an extended and improved version of a Self-Organizing Map in multiple dimensions. The goal of this methodology is to achieve high accuracy identification and impostor rejection. The proposed method, Multiple Parametric Self-Organizing Maps (M-PSOM) is a classification and verification technique. This novel method was successfully implemented and tested using the CSLU Speaker Recognition Corpora of the Oregon School of Engineering with excellent results. This method builds a unique parametric neural network for each speaker as opposed to a single neural network for the whole system as it has been done in the past. With this technology a parametric neural network is a unique representation of a speaker’s acoustic signature.

Gestures are the new silent language for controlling the human-made machines such as robotics, many recent researches toward enhancing this relation and obtaining good recognition rate were commenced, in this paper; we are trying to evaluate the performance of using multi block size for hand gesture and project this study in two cases which are the presence of edge detection operation in the preprocessing steps and the its absence. The recognition time we have achieved in case of the presence of edge detection method is 1.698 seconds and less than a second in case of none. Our suggested method includes the partitioning of the input gesture into different blocks and by using the block brightness measurement to strike the right gesture. Each gesture has been represented by a feature vector with many features controlled by the block size which in turn decides the speed and accuracy rate of the recognition. We also have modified the edge detection method in order to enhance the edges to obtain more refined and clear edges by applying a filling operation after segmentation and before edge operation. Our study also includes the application of different block sizes for generalization of the study and also can decide which block size can get good recognition rates.

Skin detection is an essential step for important vision tasks such as the detection, tracking and recognition of face, segmentation of hand for gesture analysis, and person identification. Using skin color as a feature commonly raises problem related to the varying of lighting conditions due to the complex illuminated environment in the real world. Due to this problem, many researchers describe the lighting compensation algorithm as the best solution for such a problem that is indispensable for robust skin-tone color detection. This paper presents a study on the effect of light compensation on a skin color detection model which is based on the 2-dimensional normal distribution of the skin chromatic. A number of images captured in indoor and outdoor environments are used for the experiments to prove the effect of the studied algorithm. The findings show that the lighting compensation technique gives no significant effect on the performance of parametric skin detector model and therefore this problem demands for more research work.

A new combinational window family for the design of prototype FIR filter of two-channel Quadrature Mirror Filter (QMF) bank is introduced. One variable window, viz., Kaiser Window is also used to design prototype filters. The design equations of variable window function based filter banks is also given in this article. Reconstruction error, which is used as an objective function, is minimized by optimizing the cutoff frequency of designed prototype filters. The Gradient based iterative optimization algorithm is used. The performances of filter banks designed with these window functions are compared on the basis of reconstruction error. The proposed combinational window provides the QMF bank with better reconstruction error.

The foreground detection can be utilized in tracking, segmentation or object recognition. The Local Binary Pattern (LBP) texture descriptor has been introduced for various purposes (e.g. texture classification, image indexing) and a foreground detection use is presented in this paper. Experiments on image sequences prove that the proposed algorithm, compared with other existing state of the art methods, achieves notable performance and computation speed.

Texture segmentation is the process of partitioning an image into regions with different textures containing similar group of pixels. Texture is an important spatial feature, useful for identifying object or region of interest. In texture analysis the foremost task is to extract texture features, which efficiently embody the information about the textural characteristics of the image. This can be used for the segmentation of different textured images. This paper presents a new approach for color texture segmentation using Haralick’s features extracted from color co-occurrence matrices. The originality of this approach is to select the most discriminating color texture features extracted from the color co-occurrence. Symbolic Object Approach is used for achieving texture segmentation.

This paper focuses on investigating robust edge detection methods using wavelet transform (WT) techniques. Based on the properties that the WT coefficients due to the edges concentrate and are highly correlated in the region of support (RoS) of the wavelet filters while those due to the white Gaussian noise are still Gaussian, this paper presents two robust WT-based correlation edge detectors by using the correlation of WT coefficients in the RoS of wavelet filters. The first one generates its decision statistics by integrating the point-wise energy of the WT coefficients in the RoS of wavelet filters, which is called as WT-based energy edge detector. The second one generates its decision statistics by integrating the correlation of the WT coefficients in the RoS of wavelet filters, which is called as WT-based coherent edge detector. Moreover, the statistical properties as well as the optimal integration length of these two edge detectors are derived. Finally, simulations are carried out to illustrate the detection performance of the proposed WT-based correlation edge detectors and compare them with the existing edge detection methods.

When we are encountered with a dataset constituting imprecise clusters, usually Neural Networks (NNs) are not sufficient to classify overlapped boundaries of classes. In such a situation, fuzzy processing by which vagueness is handled sufficiently may be utilized to overcome classification difficulties. In the present paper, we use an ensemble of NNs which are trained using different subsets of entire training data set. Then a fuzzy inference unit is used to process the outputs of NNs. A criterion is introduced to modify the topologies of NNs and in addition, fuzzy rules are generated simultaneously and automatically. Also a method is presented to divide the feature space into Regions of Competence (ROC). Each classifier in the ensemble will be an expert for a ROC.

The channel information capacity theorem for Shannon’s time-continuous additive white Gaussian noise channel (AWGN), widely known as the Shannon-Hartley Law, expresses the information capacity of a channel bandlimited in the conventional Fourier domain in terms of the signal-to-noise ratio in it. In this paper a modification of the Shannon’s time-continuous AWGN channel using the linear canonical transform (LCT) is presented and an expression for its capacity is derived. The Shannon-Hartley Law is shown as a special case of it. It can be observed that the capacity can be increased over and above the value predicted from the Shannon-Hartley law under some conditions. The channel capacity for AWGN channels is found to be a function of the LCT parameter.

For the last two decades the wavelet theory has been studied by many researchers to answer the demand of better and more appropriate functions to represent signals than the one offered by the Fourier analysis. Wavelets study each component of the signal on different resolutions and scales. One of the most attractive features that wavelet transformations provide is that their capability to analyze the signals which contain sharp spikes and discontinuities. Early implementations of the wavelet transform were based on filters’ convolution algorithms. This approach requires a huge amount of computational resources. In fact at each resolution, the algorithm requires the convolution of the filters used with the approximation image. Relatively recent approaches are using the Lifting Schemes (LS). In this paper we provide the taxonomy and current state of the art in Lifting Schemes (LS).

The world heritage of handwritten Arabic documents is huge however only manual indexing and retrieval techniques of the content of these documents are available. To facilitate an automatic retrieval of such hand-written Arabic document, a number of automatic recognition systems for handwritten Arabic words have been proposed. Nevertheless, these systems suer from low recognition accuracy due to the peculiarities of the handwritten Arabic language. Thus, in this Paper we propose a segmentation based recognition system for handwritten Arabic words. We divide a handwritten word into smaller pieces of a word and then these small pieces are segmented into candidate letters. These candidate letters are converted into their correspondence chain-code representation. Thereafter we extract discrete, statistical and structural features for classication. Additionally, we introduce a novel active contour based feature to increase the recognition accuracy of strongly deformed Arabic letters. We also use a decision tree to reduce the number of potential classes. We then use a neural network to compute weights for all statistical features and use them as input for a k-NN classier. Our experiments show that the extracted features by our technique achieve higher recognition accuracy as compared to other features.

Classification has emerged as a leading technique for problem solution and optimization. Classification has been used extensively in several problems domain. Automated gender classification is a significant area of research and has great potential. It offers several industrial applications in near future such as monitoring, surveillance, commercial profiling and human computer interaction. Different methods have been proposed for gender classification like gait, iris and hand shape. However, majority of techniques for gender classification are based on facial information. In this paper, a comparative study of gender classification using different techniques is presented. The major emphasis of this work is on the critical evaluation of different techniques used for gender classification. The comparative evaluation has highlighted major strengths and limitations of these existing techniques. Taking an overview of these major problems, our research is focused on summarizing the literature by highlighting its strengths and limitations. This study also presents several future research areas in the domain of gender classification.

The goal of automatic speech recognition (ASR) system is to accurately and efficiently convert a speech signal into a text message independent of the device, speaker or the environment. In general the speech signal is captured and pre-processed at front-end for feature extraction and evaluated at back-end using the Gaussian mixture hidden Markov model. In this statistical approach since the evaluation of Gaussian likelihoods dominate the total computational load, the appropriate selection of Gaussian mixtures is very important depending upon the amount of training data. As the small databases are available to train the Indian languages ASR system, the higher range of Gaussian mixtures (i.e. 64 and above), normally used for European languages, cannot be applied for them. This paper reviews the statistical framework and presents an iterative procedure to select an optimum number of Gaussian mixtures that exhibits maximum accuracy in the context of Hindi speech recognition system.