基于稀疏表示的KCCA方法及在表情识别中的应用

Kernel Canonical Correlation Analysis with Sparse Representation for Facial Expression Recognition

在面部表情识别中,由于图像特征中存在与情感语义无关的信息及噪声干扰等因素,在一定程度上影响表情识别的准确性.传统的基于核典型相关分析的识别方法难以有效克服这些因素的影响.为尽可能排除这些影响表情识别的因素,提出一种基于稀疏表示的核典型相关分析方法,并将其应用于表情识别中.该方法的基本思想是应用稀疏学习方法来自动选择表情特征矩阵中的关键特征谱成分进行表情特征与情感语义特征之间的相关性建模,然后通过建立的模型完成对待测表情图像的语义特征估计,并用于表情的分类识别.为验证所提方法较传统的基于核典型相关分析方法的优越性,选取国际标准表情数据库JAFFE进行实验,实验结果证实了所提方法的有效性.

In facial expression recognition, the existences of image noises and the irrelevant image information to the expression changes usually influence the recognition accuracy. The traditional facial expression recognition method using kernel canonical correlation analysis （KCCA） is difficulty to solve this problem. To overcome this drawback, a kernel canonical correlation analysis with sparse representation （SKCCA） is proposed and applied to the facial expression recognition. The basic idea of the SKCCA method is to utilize the sparse representation approach to choose the spectral components of the facial feature matrix before modeling the correlation between facial feature matrix and the expression semantic feature matrix. Then, the expression recognition is carried out based on the correlation model. To demonstrate the superiority of the proposed method over the traditional KCCA method, extensive experiments are conducted on the JAFFE database and the experimental results confirm the effectiveness of the proposed method.