A Research On Application Of Compressed Sensing In Pattern Classification

Candes、 Romberg、 Tao and Donoho have recently proposed the theory of Compressed Sensing which enables the reconstruction of spare or compressible signal from a non-adaptive linear measurement. If selected properly, the number of measurement can be much smaller than that needed by the Nyquist rate sample. Besides, the aim of many signal processing problems is not to recover the signal completely, but to solve the problem of pattern classification. This thesis focuses on the study of the application of the Compressed Sensing (CS) theory in the area of pattern classification.This dissertation first presents the framework of the CS theory and then focuses on the illustration of its application in the area of pattern classification from the following three perspectives:(1) the application of the CS theory in the facet of gender classification;(2) the application of the CS theory in the facet of face recognition;(3) the application of the CS theory in the detection of specific target.The main findings and contributions of this study are as follows:(1) Basing on sorting and comparing the existing methods of face detection, feature extraction and pattern classifiers, this study has succeeded in providing a integrated system of face detection and gender recognition. This system, which can be used to identify all the gender attributes appearing on the face in static pictures, consists of three parts:face detection and preprocessing, feature extraction and gender recognition. The tests of specific pictures show that this system can achieve a high recognition rate.(2) Based on the CS theory, this study has also proposed a kind of algorithm for the gender classification, which tries to solve the robust identification problem of face gender classification under complex illumination environment. The main work is to create the face gender image database, build the face gender dictionary bases and do the algorithm verification towards the subset of Extended Yale B Database, which was made in complex illumination. The results of experiments show that the computational efficiency and the recognition rate of this algorithm are higher than those of traditional algorithm.(3) This study has introduced two classification algorithms, that is, the SRC algorithm based on the sparse representation and the GSRC algorithm based on the sparse representation of Gabor features. Then the author put forward the RGSRC algorithm based on the random projection of Gabor features. This algorithm aims to study the recognition of the front and unobstructed face with relatively large changes in the expression and illumination changes. The comparative experiments are made on the AR Face Database for the verification of the algorithm and the results show that the RGSRC algorithm proposed in this thesis is better than the present other face recognition algorithms.(4) On the basis of the RGSRC algorithm, this study has also discussed the impact of measurement matrix on the accuracy of face recognition. The accuracy of face recognition with eight matrixes used as the measurement matrix respectively has been compared and analyzed. These eight matrixes are as follows:the random Gauss matrix, the random Bernoulli matrix, the random Toeplitz matrix, the random cyclic matrix, the sparse banded random Gauss matrix, the sparse banded random Bernoulli matrix, the sparse banded random Toeplitz matrix and the sparse banded random cyclic matrix. The experiments indicate that the reduction in the number of independent variables in the matrix does not exert a great influence on the recognition accuracy, which is conducive to the physical realization of the measurement matrix in the specific CS applications.(5) This study has designed a data acquisition system for the specific target’s contour. This system, which can be used to obtain the information of the specific target’s contour, is composed of five main modules:the data acquisition module, the data transmission module, the acquisition and control module, the power supply module and the processing and display module. The experimental results show that this system can collect the contour data stably and efficiently.(6) Based on the analysis of present detection algorithms, the author first discusses the particularity of the specific target’s detection and then puts forward a specific target detection algorithm which is based on the CS theory. With the use of the data acquisition system for the specific target’s contour, the contour pictures are collected and those pictures constitute a database. On the basis of this database, this algorithm can effectively complete the specific target’s detection.