Abstract: The traditional LBP feature extraction of the image is sensitive to the change of the non-monotonic light. The global feature can’t be sparsely expressed by the LBP. An adaptive weighted Local Gray Code Patterns（LGCP） and fast sparse representation of feature extraction methods is proposed. The edge detection operator is used to maximize the edge values of the original image to overcome the influence of feature description from the light changes. Eight bit gray code is got by using LGCP and is converted into decimal. The optimal representation of local features will be got by the weighted cascade block. Distribution characteristics descriptor of the cascade histogram is as the atoms to form the dictionary. The global feature of the image would have better sparse representation. Finally, a fast sparse representation is selected as a classifier for classification. Several experiments on the extended Cohn-Kanade(CK+) expression data set show that the method has a rapid recognition, and the recognition rate is up to 94%.

Key words: expression recognition, Gray Code Pattern(GCP), sparse representation

摘要： 针对传统LBP特征提取方法对非单调光线变化比较敏感且无法对全局特征进行稀疏表示的缺陷，提出一种自适应加权局部格雷码模式（Local Gray Code Patterns，LGCP）与快速稀疏表示相结合的特征提取方法。先对原始图像应用边缘检测算子最大化边缘值，以克服光线变化对特征描述的影响。采用LGCP编码得到八位格雷码并转换为十进制，然后对图像进行分块加权级联，使描述子能够对局部特征进行最优表征；同时，为了得到更好的全局特征的稀疏表示，将级联后的直方图分布特征描述子作为原子构造字典；最后，使用一种快速稀疏表示方法作为分类器进行分类识别。基于扩展Cohn-Kanade（CK+）表情数据集进行多组实验，结果表明该方法的识别速度更快，识别率可达94%。

关键词: 表情识别, 格雷码模式, 稀疏表示