Research On 2D Multi-pose Face Recognition Based On Local Invariant Feature

The intra-personal differences caused by pose variation can even be much larger than the inter-personal differences,which will lead to a decrease in the face recognition accuracy.There have been a large number of studies devoted to solving the problem of pose variation.However,pose variation is still regarded as one of the unresolved problems in the field of face recognition.As the most popular method at present,the face recognition based on view reconstruction generally has image distortion phenomenon,which would lead to misrecognition.To address the issue,this thesis aims to develop a pose-robust method,which is more natural and more consistent with the human cognitive mechanism.It means that the entire recognition process is only performed on the original face image without view reconstruction and normalization.To realize this target,this thesis focus on 2D technology and takes the local method that is less sensitive to pose variation as the starting point.Finally,two pose-robust methods are proposed: multi-pose face recognition with Huffman-LBP improved by divide-and-rule strategy and multi-pose face recognition based on hierarchical gaussian mixture model.Specific work is as follows:1.To handle the problem of losing texture information from LBP and make the sparse representation classification(SRC)be robust to pose variation,this thesis proposes a multi-pose face recognition algorithm with Huffman-LBP improved by divide-and-rule strategy.The algorithm applies Huffman coding to the feature calculation process of LBP.By the new encoding rule,the sign of the contrast value is no longer the only encoding object,the magnitude of the contrast value will also play a role in the encoding process.Therefore,the discrimination ability of the LBP can be improved.What's more,the divide-and-rule strategy is first applied to both face representation and classification with the purpose of solving the problem of pose variation.Face representation via Region Selection Factor(RSF)and Patch-based SRC fusion classification are proposed.Experiments on three multi-pose face databases(CMU PIE,FERET,and LFW)demonstrate the effectiveness of the proposed algorithm for pose variation.2.Considering that the shortcomings of patch extraction through landmarks,this thesis proposes a multi-pose face recognition algorithm based on hierarchical gaussian mixture model.First,the gaussian mixture model is utilized to find the most similar local regions in face images of different poses through unsupervised clustering.Meanwhile,in order to further optimize the effect of local region extraction,the algorithm proposes a hierarchical gaussian mixture model,which is used to select the patches in a hierarchical and progressive manner,so that the extracted patches are more similar.The results of subjective display and data comparison show that the proposed algorithm is effective.3.By summarizing the previous research experience,this thesis make a further prospect for the next research direction,that is,proposes a pose-robust face recognition framework based on hierarchical gaussian mixture model and deep learning.