基于HOSVD分类的非特定人脸表情识别算法

Nonspecific Facial Expression Recognition Algorithm Based on HOSVD Classification

由于人脸外观、光照、姿势变化等对人脸表情特征提取的影响,非特定人脸表情识别率普遍较低。针对上述问题,提出一种基于高阶奇异值分解(HOSVD)分类的非特定人脸表情识别算法。算法融合局部方向模式(LDP)全脸特征和中心化二值模式(CBP)局部特征,以增强人脸表情特征的鉴别力,引入HOSVD建立表情子空间进行分类识别,从而减少人脸外观对表情特征的影响,同时利用HOSVD求解区域能量用于精确匹配。在JAFFE数据库上的非特定人脸表情实验结果表明,HOSVD分类算法相比传统最近邻算法更能区分表情图像的特征,识别率提高了18%,此外,LDP融合CBP特征相比LDP特征和CBP特征更能准确描述人脸表情,识别率分别提高了17%和12.2%。由此可见,上述方法对解决非特定人表情识别问题具有更好的识别效果。

Due to the influence of face appearance, illumination and pose changes on facial expression feature extraction, the recognition rate of nonspecific facial expression is generally low. To solve this problem, this paper proposes a non-specific facial expression recognition algorithm based on high-order singular value decomposition(HOSVD) classification. This algorithm combines Local Directional Patterns(LDP) full face features and Centralized Binary Patterns(CBP) local features to enhance the discrimination of facial expression features. HOSVD is used to establish expression subspace for classification and recognition, so as to reduce the influence of face appearance on expression features. At the same time, HOSVD is used to calculate regional energy for accurate matching. Experimental results of non-specific facial expressions on JAFFE database show that the HOSVD classification algorithm is better than the traditional nearest neighbor algorithm to distinguish the features of facial expressions, and the recognition rate is improved by 18%. In addition, the LDP fusion CBP feature is more accurate to describe facial expressions than the LDP and CBP feature, and the recognition rate is improved by 17% and 12.2% respectively. It can be seen that this method has a better recognition effect to solve the problem of nonspecific human expression recognition.