Improved Algorithms Representation Based Classification For Face Recognition

Face recognition is of paramount importance in computer vision and biometrics systems. Recently as a result of increasing security threats led to search for effective authentication systems. Although the good performance of face recognition system, there are key robustness issues which challenge the reliability of this system. For example the variations in image configurations like pose, illumination, and facial expressions as well as occlusion and disguise remain open to challenges in the model of face recognition. To tackle these robustness problems, in these issues there is an urgent need for novel algorithms in these issues.This thesis achieves a proposed face recognition algorithms that work in general frame of sparse representation based classification (SRC) for the robustness of the main problems. In order that the model of face recognition is to tackle various problems of adverse luminance variations, severe expression variations, random pixel corruption and contiguous occlusion, three novel algorithms are proposed;First, an efficient approach for face recognition based on combination of collaborative representation based classification (CRC) and regularized least square (CRC_RLS) with bilateral filtering (BF) is proposed, where BF has shown to be an effective image denoising technique, and is better in extracting the spatial features of the image data and CRC has various instantiations by applying different norms to the coding residual and coding coefficient. This model yields considerably better performance than CRC when implemented alone. Furthermore, experiments and their results show that the proposed method outperforms several alternative methods.Second, an improved technique for face recognition consists of two phases. Initially, the Robust Principal Component Analysis (RPCA) is used specifically in the first phase, which is employed to reduce dimensionality and to extract abstract features of faces, thus a more compact and/or robust set of bases had been learned from the original images and then used as the dictionary to represent the input query image. The second phase is using metaface learned (MFL) for learning from the original images and then used it as the dictionary to represent the input query image in general framework of sparse representation based classification (SRC), in which the l1-norm minimization is more convenient for learned metafaces to be more representative and efficient. Experiments were conducted on the different databases of face images and encouraging results were shown by the experiments compared with other methods demonstrated the feasibility of this model.Third, it is an improved method, based on face recognition that renders it suitable for handling variations in image configurations. This method integrates the low-rank matrix which is recovered by using robust principal component analysis (RPCA) with relaxed collaborative representation (RCR). The purpose of the low-rank representation is that in order to obtain more discrimination information which is of benefit to face identification, and RCR, contributes to the reduction of the variance of coding vector after coding each feature vector on its associated dictionary to allow flexibility of feature coding and to address the similarity among features. Furthermore, it is characterized by the exploitation of the distinctiveness of different features by weighting its distance to other features in the coding domain. The effectiveness of the proposed methods is validated by extensive experiments on different benchmark databases and excellent results have been reported.