Study On Feature Learning For Face Recognition

Face recognition is a important issue in real world application which involving pattern recognition,artificial intelligence and computer vision area. The key point of face recognition is how to extract discriminative feature from face images to identify the specific person. Due to various adverse effects such as lighting, pose, accessory, expression and heterogeneous modality variations, the performance of most existing face recognition algorithms is not satisfactory in real world.In this thesis, we study on the local feature extract framework and present some strategies to improve the performance of face recognition algorithms. The main contributions of this thesis are:1.Study on local binary pattern and its varieties, and summarize a like local binary pattern feature extract framework. Besides, we list the pros and cons of ours framework, give the research direction of future works.2.Local Intensity Order Pattern(LIOP) is a highly discriminative descriptor which encodes the local ordinal information of each pixel.However, the discriminative power of LIOP is limited since it just describes the relationship among the four neighbors around the center point. Local Quantized Patterns(LQP) could enrich the discriminative power of local pattern. We incorporate LIOP and LQP, and proposeenhanced Local Quantized Patterns(ELQP).the experiment result indicate the proposed algorithm has better performance than compared algorithms.with WPCA, our algorithm’s performance get improved.3.To deal with the heterogeneous face recognition(HFR) problem,we propose the common discriminative subspace framework which utilizes two learning filter kernels to reduce the heterogeneous face’s difference in the same person and then apply ELQP to extract discriminative histogram features to classify each image. Unlike the traditional heterogeneous face recognition algorithm using CCA to extract features, our method can use Linear Discriminate Analysis(LDA) to utilize local the label information to enhance the algorithm’s performance.Experiment in HFB show our algorithm get good performance.