Research On Multi-View Face Recognition Based On Statistical Models

Face recognition, as an active area of research over the past two decades, still poses many challenges. Current face recognition systems yield satisfactory performance only under controlled scenarios and recognition accuracy degrades significantly when confronted with unconstrained situations due to variations such as pose, illumination and misalignments etc. After many studies, frontal face recognition has achieved good results, but multi-view face recognition is still difficult. The main object of this dissertation is to investigate and improve methods towards multi-view face recognition.Firstly, this paper described the composition of face recognition, and analyzed the various components in detail. It focused on the existing difficulties of face recognition.Secondly, the location of multi-view facial key feature points was studied in depth. By analyzing the characteristics of Gabor face image, Gabor transform was used to achieve illumination and view invariant detection of human eye windows. The method of eye location based on ASEF correlation filtering was studied deeply. ASEF correlation filter is mainly applied to the frontal face, and it assumes that the human eye in a fixed area. To solve this problem, this paper combined Gabor transform with ASEF correlation filter to locate multi-view eye in the case of unknown eye window. Experiments showed that the method of Gabor transform in combination with ASEF with high positioning accuracy in multi-view eye location.Thirdly, a view invariant face recognition method was introduced that required only frontal face images of the person to be recognized in the gallery. In the feature extraction stage, the method of extracting the local Gabor binary pattern histogram sequence was analyzed. Since the feature dimension is so high, principal component analysis was used to reduce its dimension, and linear transform was adopted to synthesize frontal facial features from non-frontal facial features, the principal components of local Gabor binary pattern histogram sequence were extracted. In face recognition stage, the proposed approach was centered on modeling joint appearance of gallery and probe images across view in a Bayesian framework. Furthermore using Gaussian density assumption was proposed to fit with the distribution between the similarities of facial features.Lastly, several experimental results were presented and compared with previous state-of-the art approaches, demonstrating the effectiveness of the proposed approach.