Abstract: Aiming at the excessive costs of the solving process of the [l1]-norm sparsity constraint on the representation coefficients, which is used in dictionary learning and classification, a new method is proposed, combining blocking LBP feature and projective dictionary pair learning to recognize the facial expression, at the meantime, to obtain more efficient facial expression-related features for dictionary learning. The blocking LBP feature of the image is extracted to replace the original data sample for training and testing. Then, this paper learns an analysis dictionary and a synthesis dictionary, and the former can obtain the representation coefficient while the latter possesses refactoring capability. Both synthesis dictionary and analysis dictionary are used to calculate the reconstruction error of each category and classify them, and facial expression recognition is achieved consequently. According to the experimental results in the JAFFE and CK+databases, compared with other methods, this proposed method not only can greatly reduce the time complexity of training and testing phases, but also can achieve higher recognition rate in classification tasks.

Key words: structured dictionary, block LBP feature, synthesis dictionary, analysis dictionary, discriminative dictionary

摘要： 针对字典学习和分类过程所采用的表示系数[l1]范数稀疏约束求解过程代价过高的问题，同时为获取更有效的表情相关特征来进行字典学习，提出一种结合分块LBP特征与投影字典对学习的表情识别方法。提取图像的分块LBP特征替代原始数据样本，用来训练和测试。学习一个分析字典和一个综合字典，分析字典可求得表示系数，综合字典具备重构能力。利用综合字典和分析字典求出各类别的重构误差进行分类，从而实现人脸表情识别。在JAFFE和CK+数据库上的实验结果表明，与其他方法相比，所提出的方法不仅可以大大降低训练和测试阶段的时间复杂度，而且可以在分类任务中达到更高的识别率。

关键词: 结构化字典, 分块LBP特征, 综合字典, 分析字典, 判别式字典