多特征多分类器优化匹配的人脸表情识别

Facial Expression Recognition Based on the Optimal Matching of Multi-feature and Multi-classifier

针对主成分分析(Principal Component Analysis,PCA)降维过程中由于特征值相对集中而造成维数仍然偏高的不足,本文提出了基于最优样本的主成分分析(Optimal Sample-PCA,OS-PCA)降维方法。OS-PCA通过选择训练样本、优化协方差矩阵,从而达到进一步降维的目的。鉴于离散余弦变换(Discrete Cosine Transform,DCT)对光照的鲁棒性,以及局部二值模式(Local Binary Patterns,LBP)对局部纹理特征的有效描述,本文结合DCT和LBP特征来弥补单一OS-PCA特征在人脸表情表征方面的局限性。为了更好地发挥特征与分类器的协作优势,文章构造了一个三层多分类器最优集成的人脸表情识别模型。该模型首先对表情图像进行预处理操作;然后提取OS-PCA、DCT和LBP特征送入三层模型;最后基于单一特征和相应单一分类器的最佳匹配组合,完成多特征与多分类器的最优集成;在执行粗分类结果投票表决的基础上,进一步对仍有分歧的表情图像进行自适应决策,从而得到最终识别结果。实验表明,OS-PCA较PCA进一步有效地降低了特征维数;同时,基于多特征多分类器的三层识别模型在JAFFE和CK库上分别获得了高于95%和96%的识别率,并表现出比较优越的时间性能。

Principal Component Analysis（PCA） can effectively extract global features from images and has advantages of dimension reduction. During the dimension reduction process, because of the comparatively concentration of eigenvalues, the dimension is still larger than the best. To solve this problem, this paper presents the optimal-sample PCA（OS-PCA） for dimension reduction. By choosing the training samples and optimizing the covariance matrix, OS-PCA achieves the purpose of further dimension reduction. Because Discrete Cosine Transform（DCT） has robustness of light, as well as Local Binary Pattern（LBP） is effective in describing local texture features, the paper combines DCT and LBP features to make up for the limitations of OS-PCA in facial expression representation. In order to utilize the advantages of collaboration features and classifiers, this paper constructs a facial expression recognition model, which is based on three layers of the optimal integration of multiple classifiers. Firstly, facial images are preprocessed. This step includes the detection of face from images and normalization. Then the OS-PCA, DCT and LBP features are delivered into the model. Finally, based on the best match combination between single classifier and single feature, the model completes the optimal integration of multiple features and multiple classifiers. Via voting mechanism, the model makes adaptive decisions for images that are still different to get the final recognition result. Experiments show that OS-PCA is more effective than PCA in dimension reduction. On the JAFFE and CK database, recognition rates are higher than 95% and 96%, and the proposed model shows brilliant time performance.