Research On Face Recognition Based On Sparse Representation

As a kind of biologic character, human face is portable and is never lost or stolen, and the way to capture face image is simple and friendly without cooperation, even the capturing can not be perceived. So face recognition, as a biometric identification, has been extensively studied in the fields of pattern recognition and computer vision, Face recognition technology will be used in many areas such as human-computer interaction, security, law enforcement, entertainment and so on in future. However, the performances of face recognition approaches are influenced by expression, illumination, occlusion and pose, and there are some intractable problems will need to be further gone into.Sparse representation is highly competent for classification and it has been studied widely in the field of pattern recognition. Construction of over-completed dictionary and solution of sparse representation coefficient vector are two main problems in sparse representation. In this paper, our research focuses on dictionary learning with all training face images, fast solution of sparse representation and promotion of the recognition accuracy for face recognition based on sparse representation. Researches done in this paper were summarized as follows:(1) Face recognition using plurality voting and collaborative representation based on binary bit-plane images. The theories about sparse representation and collaborative representation were investigated in this paper. To improve the recognition accuracy of approaches based on collaborative representation and the recognition speed of approaches based on SRC, bit-plane images and plurality voting were integrated into collaborative representation.To enhance contour, original gray face images were first equalized by cumulative distribution function and then were decomposed into eight different binary bit-plane images for the image with 256 gray levels used in this paper. Those bit-plane images having the same bit order constructed the same one bit-plane image database. The accuracy of each database was obtained using the collaboration representation classification on the training stage. Training results showed that the 1st,5th,6th,7th and 8th bit-planes containing more identifiable information were selected to take part in recognition. For the five recognition results, plurality vote was used to decide the only proper identity of the testing. If plurality voting results were not unique, the voting fails and the second decision should be executed. Collaboration representation classification performed again on the virtual weighted face database, and the virtual weighted face images were constructed by all bit-plane images. Weight vector was determined by the bit recognition accuracy rate and the bit order. The bit-plane images extracted from the gray face images are binary, but are not gray images with two levels. The proposed bit-plane extraction approach could avoid lower-bit-plane images being immersed in higher-bit-plane images in the virtual weighted image. Those bit-planes with less recognition information and the lower bit-plane whose accuracy is the same to the higher-bit-plane were all discarded for plurality voting. Weighted vector was related to the recognition accuracy and the order of each bit-plane, which embodies that each bit-plane has different contribution to the virtual images and ensures that the differences between them are limited. The accuracy of this proposed approach reached 97% and 98% on ORL and FERET database respectively. Bit-plane decomposition, weighted vector training and construction of virtual weight training image had been completed on the training stage, so recognition speeds of this proposed approach and CRC_RLS are comparative, which are more than ten times that of SRC approach.(2) Collaborative Representation based Classification for face recognition with neighbors selected in LCP feature space. The local configuration pattern (LCP) feature includes local structural information and microscopic configuration information. The local structural information represented by local binary pattern (LBP) contains low-level properties and basic visual elements, and it can efficiently detect the local structures such as edges and lighting spots. The microscopic configuration information represented by microscopic configuration model (MiC) embodies image configuration and pixel-wise interaction relationships using a linear model. In LBP feature space, χ2-LBP-similarity and χ2-BRD-LBP-similarity were defined respectively based on normal histogram and on the bin-ratio histogram utilizing negative logarithm of Chi square coefficient to evaluate the similarities between images. The similarity measurements proposed in this paper are competent to indicate the differences between histograms. Euclidean distance was employed to estimate the similarity between images in MiC feature space. The value ranges both of validity threshold and neighboring threshold for the normal histograms and for bin-ratio histograms, and the value ranges both of validity and neighbor threshold for the MiC feature were all provided experimentally in this paper. Through intersecting or unionizing the similarities estimated in both LBP level and MiC level, neighbors of testing were chosen and as column vectors to compose the over-completed compact dictionary adaptively. So dictionary atoms are more similar to the testing image, which is beneficial to accuracy. The amount of atoms is variable and is about 2/3 of the overall number of training images for each testing image, which significantly reduces the computational cost in coding the testing sample. The proposed approach using the dictionary constructed based on the bin-ration histogram is more robustness of occlusion. Compared with CRC_RLS, the accuracy of this proposed approach was increased by about 3% without occlusion. And the accuracies with the occlusion of sunglasses and scarf were improved greatly and reached 85%.Face recognition is a classic small-size-sample problem, and the dictionary directly composed by face images is not satisfied with the sparse representation requirement that the number of atoms is much more than that of atom-features, so PCA is employed to reduce dimension before classification with CRC RLS.(3) Face Recognition using compressive sensing for dimension reduction and collaborative representation by neighbors selected in Gabor feature space. Face images were first mapped into the low-dimension Gabor feature space via down-sampling Gabor features. Then Chi square coefficient was employed as the similarity measurement to estimate the similarity among images. Representation matrix is composed adaptively with the neighbors of testing image that selected by class-average-correlation-coefficient from each class. So representation bases are more similar to the testing image and they distributed over all classes, which contributes to correctly classify. In addition, the number of representation bases is about half of the number of all training images, which decreases the computation cost of solution to sparse representation vector. As sensing matrix, Gaussian random matrix sensed face images and reflected the very high dimension face images into any low-dimension measurement space, which overcomes the troublesome issues brought by the feature extraction and selection which is inevitable for the feature subspace approaches. Compressive sensing reduction dimension got rid of the restriction that sample-feature number must be smaller than sample number in PCA and the feature number is arbitrary, which is a good solution to the small-size-sample problem.Compared with SRC, recognition speed has been increased five times and the accuracy has been increased by 5% without occlusion. For both sunglasses occlusion and scarf occlusion, the accuracies reach 73% and 83% respectively on AR database.