Features Selection And Axtraction Of Face Recognition With Occlusion

Face recognition with occlusion is a critical problem for real-world face recognition system.The main difficulty lies in the feature losing and the local aliasing caused by occlusions.For a long time,a popular viewpoint considers that feature extraction cannot deal with facial occlusion effectively.However,recent studies shew that feature selection and extraction are not only necessary,but also very important to handle facial occlusions without explicitly representing and detecting occlusion.In this work,we therefore mainly study feature selection and extraction methods for face recognition with occlusion,and we also emphasize their importance.The main research contents of this paper mainly include the following two aspects:(1)In order to effectively separate occlusion from facial images,and select the useful features,we proposed an SVD-based Gabor occlusion dictionary(GOD)learning method.We explored the covariate shift problem caused by occlusion and illumination changes from the view of dictionary coding.We noted that the original K-SVD-based GOD learning method is very time-consuming,highly redundant and lack of natural occlusion structure.The proposed SVD-based GOD learning method rectified these shortages.The advantage of the SVD method are further verified by comparing the three computational methods: K-means,K-SVD and SVD.Experiment shew that the classification method based on the SVD-based GOD learning has a much better recognition performance than the classification method based on the K-SVD-based GOD learning method in various situations.(2)An occlusion-robust feature representation method,called Adaptive Weberfaces(AdapWeber),is proposed.Based on the two guiding principles,redundancy and spatial locality,indicated by Wright et al,we analyzed the two Weber ratios implied in Weber's law and shew that the two Weber ratios happen to take charge of the two guiding principles,respectively.Based on the two Weber ratios,we proposed the AdapWeber.The facial features produced by AdapWeber are both spatially local and redundant.In order to further improve the redundancy and spatial locality of AdapWeber,we extended AdapWeber to its Multi-Scale and Multi-Orientation variant,and proposed a Multi-Scale and Multi-Orientation AdapWeber(MSMO-AdapWeber).Experiments shew that the features extracted by MSMO-AdapWeber are robust to occlusions,especially when the occlusion level is very high or the image dimension is very low.The two proposed methods and the relevant popular methods are tested and compared on the following three publicly available benchmark databases: Extended Yale B,AR and UMB-DB.Experimental results demonstrated the efficacy of the two proposed methods.Although the two methods still belong to the classical manually designing methods,they also have many performance advantages compared to the recently popular deep learning methods,such as PCANet.In the future research work,we will focus on the occlusion robustness of the deep learning methods,such as the convolutional neural networks.