关键词: 邻近二值模式关系, 多区域特征融合, D-S证据理论, 可信度修正

Abstract: To extract the regional features which are conducive to expression classification and achieve the decision-level fusion of multi-regional features, this paper proposes an expression recognition frame based on neighbor binary pattern relation (NBPR) feature descriptor and credibility modification evidence fusion (CMEF). Firstly, for the traditional local binary pattern (LBP) operator only takes the fixed center pixel as threshold, NBPR is proposed, which can encode the XOR relationship of binary pattern between adjacent pixels from multi-orientations in a local region. Then, the extracted NBPR texture features of three salient regions such as eyebrows, eyes and mouth are used to construct the initial probability assignments of evidences. Finally, in view of the deficiency of conflicting evidences combination with Dempster-Shafer (D-S) theory of evidence, a new combination method based on the evidence credibility is used to achieve the decision fusion of the three regional evidences. Experiments of the proposed method are performed on Cohn-Kanade (CK), it achieves an average expression recognition rate of 94.67% and an average time of recognition of 752 ms. The experimental results show that the NBPR is conducive to texture descriptor and CMEF strategy is beneficial to decision-level fusion, so the proposed method has higher expression recognition rate.

Key words: neighbor binary pattern relation (NBPR), multi-regional feature fusion, Dempster-Shafer theory of evidence, credibility modification