Research On Face Recognition Under Pose Variations And Difficult Lighting Conditions

Face recognition is a significant issue in the area of pattern recognition, computer vision, etc. In the past few decades, face recognition under frontal views has made satisfactory progress. However, the recognition rates for face images under pose variations and difficult lighting conditions still need to be improved. This work focuses on face recognition under pose variations and difficult lighting conditions, and the contributions are as follows:1. A dictionary learning method, Bilinear Discriminative Dictionary Learning (BDDL), is proposed for face recognition under pose and illumination variations. The proposed BDDL has the following characteristics:1) theι2-norm regularization is imposed on the coding coefficients, which makes the learned dictionary more suitable for face recognition.2) By minimizing the reconstruction error for each class using the dictionary atoms associated with that class, we learn a structured dictionary which is able to make the reconstruction error for each class more discriminative for classification.3) To exploit the discriminability of the coding coefficients of samples coded over the learned dictionary, a discriminative term bilinear to the training samples and the coding coefficients is incorporated in BDDL. The bilinear discriminative term essentially resolves a regression problem in the Reproducing Kernel Hilbert Space (RKHS). Additionally, a novel classifier, the Bilinear Discriminative Classifier (BDC), is proposed in this paper. Based on the K-L divergence, a criterion is proposed to evaluate the confidence of BDC. Experimental results demonstrate that the proposed dictionary learning method is effective to face recognition under pose and illumination variations.2. Local Histogram Specification using Learned histograms (LHS-L) and flexible LHS-L (f-LHS-L) are proposed to preprocess face images under difficult lighting conditions. Firstly, a high-pass filter is applied on face images to filter the low frequency illumination. Subsequently, local histograms and local histogram statistics are learned from the normal lighting images. In LHS-L, Local Histogram Specification (LHS) is applied on the entire image. By contrast, in f-LHS-L, the regions contain high frequency illumination and weak face features on the test image are identified by the local histogram statistics, before LHS is applied on these regions to eliminate high frequency illumination and enhance weak face features. In addition, an algorithm is also proposed to efficiently calculate the cumulative distribution function value for a particular gray level during each iteration of local histogram specification/equalization methods, and the efficiency is theoretically guaranteed.3. The classifier ensemble BDDL (En-BDDL) is proposed to cope with face recognition under both pose and illumination variations. For illumination variations, an illumination preprocessing method is adopted to eliminate illumination variations. Since BDDL is able to deal with pose variations and its learning process is based on optimizing three classifiers: the collaborative representation based classifiers (CRC_RLS and CRC_LS) and the BDC, we utilize BDDL to learn two dictionaries for the two views of images, the non-preprocessed images and the preprocessed images, respectively. In the classification process, with the confidence of BDC taken into consideration, the results of six classifiers in the two views are assembled to make the final decision.