Face Recognition Algorithm Based On Compressive Sensing

As a classic problem of pattern recognition, face recognition involves facedetection, pretreatment, feature extraction and classification. It is the integration ofpsychology, pattern recognition, computer vision science. Face recognition hasreceived the widespread attention for face image is easy to collect and doesn’t needthe person’s special match to complete identification work. To raise the accuracy offace recognition, the face recognition supports more research on the facial expression,pose, illumination, gender and disguise and so on.The proposing and development of compressive sensing brings new inspiration tofeature extraction of face recognition and make the face recognition technology basedon sparse representation widely researched. The dissertation has done some researchwork about face recognition based on compressive sensing and puts forward someimprovement, and the main work is as follows: a face recognition method based onwavelet packet transform(WPT) and compressive sensing is proposed. Firstly, itdecomposes the face image into four sub-bands using wavelet packettransform(WPT),then obtains the composition of sparse matrix dictionary. Secondly,the algorithm reduces the dimension and extracts effective information by projectionmatrix to gain the final eigenvector. Experiment result shows that the algorithm hashigher recognition rate and good robustness with face expression, posture anddisguise. Without conducting reconstruction algorithm to compute the optimal sparsesolution, it greatly reduces the computational complexity of compressive sensing. Amethod of choosing the optimal wavelet basis based on the minimum reconstructionerror is proposed. Different wavelet basis is suitable for different image to gainfeature extraction. To choose the optimal wavelet basis for face images, the choicecriteria is put forward. It also proposes a feature extraction algorithm based onquaternion and compressive sensing. The algorithm has the versatility, effectivenessand strong extensibility. It overcomes the previous algorithm limitations only fit forthe one type of image feature extraction.