Improved subspace learning for facial image analysis

73 p.Face and facial expression recognition research has been motivated by wide and potential applications. Existing facial image analysis usually suffers from the wellknown curse of imensionality and small sample size problems. In this project, our contributions are two-folds. The first is to make a comparative study on conventional subspace learning algorithms for facial image analysis (including face and facial expression recognition). Linear subspace learning has been extended to bilinear subspace learning throughout the development recently. Thus, it is useful to make a detailed study on the subspace learning considering both the linear and bilinear approaches. Another contribution is on analyzing and studying the effect of misalignment for facial image analysis. Misalignment is another issue which heavily affects face and facial expression recognition performance. Our contribution in this project is a misalignment-robust subspace learning method to improve the recognition performance of existing subspace learning algorithms for misaligned facial images. First, we conduct an experiment to investigate the influence of three misalignment cases, such as translation, rotation and scale for face and facial expression recognition. Then, we present a misalignment-robust subspace learning algorithm to improve the robustness of existing misaligned facial image analysis. We randomly generate virtual samples with the combination of the three variation cases to learn a robust subspace for feature extraction. Results of these experiments are discussed in this reportMaster of Science (Signal Processing