Abstract: Facial expression recognition, one of the most important approaches to delivery emotion, attracts many researchers?? attention in these years as a crucial part of affective computing. As similar to classical pattern recognition tasks, the key of expression recognition is how to represent more discriminative statistical features. Meanwhile, how to combine different types of features efficiently plays a vital role in performance improvements, since various kinds of features have weighted contributions. In this paper, mark points are selected according to significant geometric changes to form geometric features, and appearance features are extracted from blocks constructed based on geometric feature. Multi-kernel trick is also employed in the classifier. Experiments on a public dataset (CK+) and a self-build dataset (JNFE), compared with classical methods and state-of-the-art methods, 96.90% and 92.85% accuracy rates are achieved respectively, proving the efficiency of proposal method.

Key words: facial expression recognition, texture features, geometric features, multi-kernel fusion

摘要： 人脸表情作为人类情感的重要传达方式，近年来作为情感计算的重要组成部分，人脸表情识别吸引了很多学者的关注。与其他模式识别的问题类似，构造更为有效的统计特征是解决人脸表情识别的关键所在。同时，由于不同种类的特征对于模型性能的贡献不同，有效地利用不同特征对于性能的提升也至关重要。根据几何显著变化筛选标注点以形成几何特征，并根据几何特征构造特定的块形成形态特征；后采用多核多特征融合方法进行表情识别。通过在公开数据集（CK+）和自建数据集（JNFE）上的实验，和一些视频序列表情识别方法对比，分别获得了96.90%和92.85%的准确率，证明了所提方法的有效性。

关键词: 面部表情识别, 纹理特征, 几何特征, 多核融合