The Design Of New Classifiers And Decision Fusion Algorithms For Face Recognition

Face recognition is the hot research topic in the field of pattern recognition and computer vision. Classifier design is the key problem in face recognition and its performance plays a great role in the entire system. There are certain problems with the traditional face recognition classifiers:complexity, unsatisfactory recognition rates and weak performance under conditions of varying facial expression, illumination and occlusion.This paper addresses these problems and conducts a comprehensive investigation into classifier design. Some face recognition classifiers are developed.The least squares classifier represent the feature vector of a test sample as a linear combination of samples from a single class, then parameter vector is solved by least squares method. Recognition is determined on the estimate errors with respect to different classes. This approach is simple and fairly effective; can be easily realized.Sparse representation has been widely applied to many fields and became a hot research topic in face recognition. This paper performs a comprehensive investigation into the sparse representation theory and the design of sparse representation classifier. In sparse representation based classification; all the training samples are used to construct a over-complete dictionary and sparse representation coefficients of all testing samples are recovered by the same dictionary. The large dictionary size results in great computation consumption; moreover; the differences between testing samples are ignored. This paper addresses these problems and proposes a novel least squares algorithm to learn an adaptive dictionary for each probe sample. The learned dictionary is quite different from the unique and permanent dictionary in sparse representation based classification; its size is greatly reduced. It leads to improvement in computational efficiency and robustness.Partition scheme is observed to be a useful approach to improve recognition performance under conditions of varying facial expression, illumination and occlusion. However; each partition has different quality and they are recognized with different confidences. The traditional decision fusion algorithms such as majority voting treats the recognition result of each partition equally and the confidence differences in the partition recognition results are ignored. In this paper; a confidence index based decision fusion algorithm has been developed. Our approach can distinguish the recognition confidence of each partition and all partitions are involved in the decision making process. The non-face partitions weights less in decision fusion and complementary recognition information of face partition is fully used to achieve the final decision. Thus; performance under conditions of varying facial expression, illumination and occlusion can be improved.