Abstract: Aiming at the shortcomings of Directional Local Binary Pattern（DLBP） that can obtain the feature of image texture only at single scale, an expression recognition algorithm of Asymmetric Region-Directional Local Binary Pattern（AR-DLBP） multi-scale multi-directional fusion is proposed. Firstly, the face expression image is subjected to light compensation preprocessing to eliminate the influence of light and noise, and the key region of the face and eyebrows, eye and mouth are divided, and the contribution degree（CM） of the key area is calculated. It  extracts histogram feature information of the XOR-Asymmetric Region-Directional Local Binary Pattern（XOR-AR-DLBP） of face and critical areas, and weights the cascade of the critical region histogram information and the feature of the whole face image information fusion. Finally, it uses SVM classifier for expression classification and identification. The algorithm has the average recognition rate of 95.71% and 97.99% and the average recognition time of 112 ms and 135 ms respectively in the JAFFE library and CK library. The experimental results show that the algorithm can effectively and accurately complete the recognition of facial expression. The algorithm prepares and divides the critical region of expression by weight compensation and expresses the local and global features, which greatly improves the recognition ability of the feature and contrasts with the traditional algorithm. It is also shown that the algorithm, both in the recognition rate and recognition time, the results are the best.

Key words: facial expression recognition, Asymmetric Region-Directional Local Binary Pattern（AR-DLBP）, multi-feature fusion, Support Vector Machine（SVM）

摘要： 针对方向性局部二值模式（DLBP）在单尺度下获取图像纹理特征的不足，提出一种非对称方向性局部二值模式（AR-DLBP）多尺度多方向融合的表情识别算法。首先对人脸表情图像进行光照补偿预处理，消除光照、噪声的影响，分割出人脸及眉、眼、嘴局部表情关键区域，并计算出关键区域的贡献度（CM）；然后提取人脸及关键区域的异或-非对称方向性局部二值模式（XOR-AR-DLBP）直方图特征信息，并根据CM对关键区域直方图信息进行加权级联再与整幅人脸图像的特征信息进行融合；最后用SVM分类器进行表情分类识别。该算法在JAFFE库、CK库上仿真实验，分别取得95.71%、97.99%的平均识别率及112?ms、135?ms的平均识别时间，实验结果表明，该算法可以有效精确地完成人脸表情的分类识别。通过对表情图像光照补偿预处理及分割出表情的关键区域，并加权融合局部与整体特征，大大提高了特征的鉴别能力，与传统算法的对比实验，也表明该算法无论是在识别率还是在识别时间上，所得效果都是最好的。

关键词: 表情识别, 非对称方向性局部二值模式, 多特征融合, 支持向量机（SVM）