Seyed Mostafa Mirhassani

Self-Adjustable Neural Network for speech recognition

Neural networks with fixed input length are not able to train and test data with variable lengths in one network size. This issue is very crucial when the neural networks need to deal with signals of variable lengths, such as speech. Though various methods have been proposed in segmentation and feature extraction to deal with variable lengths of the data, the size of the input data to the neural networks still has to be fixed. A novel Self-Adjustable Neural Network (SANN) is presented in this paper, to enable the network to adjust itself according to different data input sizes. The proposed method is applied to the speech recognition of Malay vowels and TIMIT isolated words. SANN is benchmarked against the standard and state-of-the-art recogniser, Hidden Markov Model (HMM). The results showed that SANN was better than HMM in recognizing the Malay vowels. However, HMM outperformed SANN in recognising the TIMIT isolated words.

Fuzzy-based discriminative feature representation for children's speech recognition

Article history: Available online 9M ay 2014 Automatic recognition of the speech of children is a challenging topic in computer-based speech recognition systems. Conventional feature extraction method namely Mel-frequency cepstral coefficient (MFCC) is not efficient for childrens speech recognition. This paper proposes a novel fuzzy-based discriminative feature representation to address the recognition of Malay vowels uttered by children. Considering the age-dependent variational acoustical speech parameters, performance of the automatic speech recognition (ASR) systems degrades in recognition of childrens speech. To solve this problem, this study addresses representation of relevant and discriminative features for childrens speech recognition. The addressed methods include extraction of MFCC with narrower filter bank followed by a fuzzy- based feature selection method. The proposed feature selection provides relevant, discriminative, and complementary features. For this purpose, conflicting objective functions for measuring the goodness of the features have to be fulfilled. To this end, fuzzy formulation of the problem and fuzzy aggregation of the objectives are used to address uncertainties involved with the problem. The proposed method can diminish the dimensionality without compromising the speech recognition r ate. To assess the capability of the proposed method, the study analyzed six Malay vowels from the recording of 360 children, ages 7 to 12. Upon extracting the features, two well-known classification methods, namely, MLP and HMM, were employed for the speech recognition task. Optimal parameter adjustment was performed for each classifier to adapt them for the experiments. The experiments were conducted based on a speaker-independent manner. The proposed method performed better than the conventional MFCC and a number of conventional feature selection methods in the children speech recognition task. The fuzzy-based feature selection allowed the flexible selection of the MFCCs with the best discriminative ability to enhance the difference between the vowel classes.

Improvement of Hessian based vessel segmentation using two stage threshold and morphological image recovering

In many of vessel segmentation methods, Hessian based vessel enhancement filter as an efficient step is employed. In this paper, for segmentation of vessels, HBVF method is the first step of the algorithm. Afterward, to remove non-vessels from image, a high level threshold is applied to the filtered image. Since, as a result of threshold some of weak vessels are removed, recovering of vessels using Hough transform and morphological operations is accomplished. Then, the yielded image is combined with a version of vesselness filtered image which is converted to a binary image using a low level threshold. As a consequence of image combination, most of vessels are detected. In the final step, to reduce the false positives, fine particles are removed from the result according to their size. Experiments indicate the promising results which demonstrate the efficiency of the proposed algorithm.

Classification of remote sensing images from urban areas using Laplacian image and Bayesian theory

This paper presents the methodology of urban area classification in high resolution satellite IKONOS imagery. The strategies include building extraction by Bayesian theory and laplacian criterion, labeling and size filtering, intensity threshold and etc which are applied to IKONOS image in tandem to make this algorithm as an effective strategy to save processing time and improve robustness. To realize the strategy, First, vegetation are extracted in attend to green layer of RGB image then buildings are detected by Bayesian decision theory in regard to laplacian probability density function, then shadows which have low intensity are detected. In the next step a special intensity level is calculated as a threshold level to discern roads. Finally open areas are extracted from remained of image as regions with low laplacian intensity and large size. Meanwhile morphological operations are applied to remove redundant images particles. Experimental result indicates that this approach has a high efficiency especially in extraction of large roads and streets from dense urban area IKNOS images.

Hierarchical segmentation of urban satellite imagery

Abstract This paper proposes a method to combine contextual, structural, and spectral information for classification. The method is an integrated method for automatically classifying urban-area objects in very high-resolution satellite imagery. The approach addresses three aspects. First, the Gabor wavelet is applied to the image along with morphological operations, with the sparsity of the outcome considered. A Bayesian classifier then categorizes the different classes, such as buildings, roads, open areas, and shadows. There are some false positives (wrong classification), and false negatives (non-classification) in the initial results. These results can be corrected by the relaxation labeling categorization of the unknown regions. The novelty of the proposed approach lies in the extensive use of spatiotemporal features considering the sparsity of urban objects. The results indicate improvement in classification through relaxation labeling compared with existing methods.

Hierarchical method for building extraction in urban area’s images using unsharp masking [USM] and Bayesian classifier

Recently, due to the availability of high resolution IKONOS image, classification of remote sensing images from urban area become one of the most attractive topics for scientific researches and papers. In this paper, we address a method for classification of remote sensing (IKONOS) image and especially for the extraction of buildings. First step is applying the unsharp masking [USM] which intensify high frequency components of the original image. Then imagepsilas Laplacian, Bayesian classifier and size filter is used for building discrimination. The accuracy of small and large building classification using unsharp mask filter and Bayesian discrimination function is increased compared with the original qualitative model for Bayesian classification. Experiments indicate promising results about the efficiency of the proposed approach.

Fuzzy based foreground background discrimination for probabilistic color based object tracking

In the most of object tracking tasks dealing with partial occlusion is a challenging issue. Recently the use of color cue based on Monte Carlo tracking method and particle filtering is mostly considered to overcome the problem of partial occlusion and nonrigid motion. The proposed approach in this paper is based on using of sequential Monte Carlo and particle filtering for tracking. But in this method a special fuzzy based color model for object is employed. Then comparison of mean value of reference and candidate window in the proposed color space is utilized for tracking of objects. Some of the morphological operation is also used to provide a unit region for object location in the fuzzy based color space. Experimental results indicate that the algorithm is efficient in dealing with partial occlusion.

Age Estimation Based on Children’s Voice: A Fuzzy-Based Decision Fusion Strategy

Automatic estimation of a speakers age is a challenging research topic in the area of speech analysis. In this paper, a novel approach to estimate a speakers age is presented. The method features a “divide and conquer” strategy wherein the speech data are divided into six groups based on the vowel classes. There are two reasons behind this strategy. First, reduction in the complicated distribution of the processing data improves the classifiers learning performance. Second, different vowel classes contain complementary information for age estimation. Mel-frequency cepstral coefficients are computed for each group and single layer feed-forward neural networks based on self-adaptive extreme learning machine are applied to the features to make a primary decision. Subsequently, fuzzy data fusion is employed to provide an overall decision by aggregating the classifiers outputs. The results are then compared with a number of state-of-the-art age estimation methods. Experiments conducted based on six age groups including children aged between 7 and 12 years revealed that fuzzy fusion of the classifiers outputs resulted in considerable improvement of up to 53.33% in age estimation accuracy. Moreover, the fuzzy fusion of decisions aggregated the complementary information of a speakers age from various speech sources.

Gender Classification in Children Based on Speech Characteristics: Using Fundamental and Formant Frequencies of Malay Vowels

Speech is one of the prevalent communication mediums for humans. Identifying the gender of a child speaker based on his/her speech is crucial in telecommunication and speech therapy. This article investigates the use of fundamental and formant frequencies from sustained vowel phonation to distinguish the gender of Malay children aged between 7 and 12 years. The Euclidean minimum distance and multilayer perceptron were used to classify the gender of 360 Malay children based on different combinations of fundamental and formant frequencies (F0, F1, F2, and F3). The Euclidean minimum distance with normalized frequency data achieved a classification accuracy of 79.44%, which was higher than that of the nonnormalized frequency data. Age-dependent modeling was used to improve the accuracy of gender classification. The Euclidean distance method obtained 84.17% based on the optimal classification accuracy for all age groups. The accuracy was further increased to 99.81% using multilayer perceptron based on mel-frequency cepstral coefficients.

A magnetic resonance imaging study on the articulatory and acoustic speech parameters of Malay vowels

The phonetic properties of six Malay vowels are investigated using magnetic resonance imaging (MRI) to visualize the vocal tract in order to obtain dynamic articulatory parameters during speech production. To resolve image blurring due to the tongue movement during the scanning process, a method based on active contour extraction is used to track tongue contours. The proposed method efficiently tracks tongue contours despite the partial blurring of MRI images. Consequently, the articulatory parameters that are effectively measured as tongue movement is observed, and the specific shape of the tongue and its position for all six uttered Malay vowels are determined.Speech rehabilitation procedure demands some kind of visual perceivable prototype of speech articulation. To investigate the validity of the measured articulatory parameters based on acoustic theory of speech production, an acoustic analysis based on the uttered vowels by subjects has been performed. As the acoustic speech and articulatory parameters of uttered speech were examined, a correlation between formant frequencies and articulatory parameters was observed. The experiments reported a positive correlation between the constriction location of the tongue body and the first formant frequency, as well as a negative correlation between the constriction location of the tongue tip and the second formant frequency. The results demonstrate that the proposed method is an effective tool for the dynamic study of speech production.