Study Of Feature Extraction Approaches For Color Face Images

With the scientific and technological development, face recognition has gradually become an important research topic of identity validation techniques, and color face recognition has attracted more and more attention. Color images afford more information than gray-scale images. How to fully utilize the complementary information between color components, reduce their redundancy and extract effective features is a key problem of color face recognition. To solve this problem, this dissertation focuses on feature extraction of color face images. The main research content and innovative achievements are as follows:(1) Two novel feature extraction approaches, called similarity preserving analysis based on sparse representation(SPASR) and sparsity preserving embedding with manifold learning and discriminant analysis(SPE-MD), are proposed for color face image recognition. They select three components with the minimum similarity from the commonly used RGB, HSV, YCbCr and YIQ spaces by using a simple color component selection strategy, and then fuse the selected components on the data level. For the fused face image samples, SPASR seeks a low-dimensional feature space in an unsupervised manner, which can preserve the sparse reconstructive similarity relations among samples, and also maintain the similarities of original samples and sparsely reconstructed samples; the supervised SPE-MD approach aims at both preserving the sparse reconstructive relations among original samples and maintaining the manifold and discriminant information of original samples and sparsely reconstructed samples. These two approaches can effectively reduce the data-level similarity between three color components, and make full use of the sparse reconstructive relations among data-level-fused image samples to extract favorable face features. The public AR, Face Recognition Grand Challenge version 2 and Labeled Faces in the Wilds color face image databases are used as test data. Experimental results demonstrate that the recognition performance of SPASR is better than that of three representative unsupervised data-level-fused color face image feature extraction and recognition methods, and SPE-MD achieves better recognition results in comparison with SPASR and four representative supervised data-level-fused color face image feature extraction and recognition methods.(2) Two new discriminant feature extraction approaches, named holistic orthogonal analysis(HOA) and statistically orthogonal analysis(SOA), are presented for color face image recognition. They employ the discriminant analysis within each color component and unsupervised feature-level similarity analysis between different components. HOA and SOA in turn extract the features of R, G and B components by using the Fisher criterion, simultaneously separately make the achieved three projective transformations mutually orthogonal and statistically orthogonal to remove the similarity between features extracted from different components, and then fuse the features from three components to do recognition. These two approaches can remove the feature-level similarity between three color components of face images, and take full advantage of the discriminant information within each component and complementary information between different components to obtain effective discriminant features. Recognition results on the three color face image databases demonstrate that HOA and SOA perform better than SPASR, SPE-MD and two representative feature-level-fused color face image feature extraction and recognition methods.(3) Two novel discriminant feature extraction approaches, named Euclidean distance metric based dual discriminating similarity analysis(EDM-DDSA) and correlation metric based dual discriminating similarity analysis(CM-DDSA), are proposed for color face image recognition. They employ the dual feature-level discriminating similarity analysis of color components. EDM-DDSA and CM-DDSA separately employ the Euclidean distance and correlation metrics to perform feature-level discriminating similarity analysis both within each color component and between different components. They preserve the favorable features’ similarity between same-class samples and remove the adverse features’ similarity between different-class samples while extracting discriminant features from three components, and then fuse the features from three components to recognize faces. Experimental results on the three color face image databases demonstrate that EDM-DDSA and CM-DDSA further improve the recognition results as compared with HOA and SOA.