基于非凸低秩分解双字典误差模型的遮挡表情识别

摘要

In order to effectively overcome the error of expression recognition caused by occlusion,and to reduce the dependence of ex-pression recognition on identifying individuals,a novel occlusion expression recognition method based on non-convex low rank decom-position double dictionary error model is proposed in this paper. Firstly,to solve the problem that the kernel norm approximation rank function can not effectively estimate the rank of the matrix,the non-convex logarithm function to approximate the rank function is pro-posed,which can improve the estimation accuracy and sensitivity to noise. The expression features and identity features in each type of expression image are separated by non-convex logarithmic function low rank decomposition,and then two part features are respectively dictionary-learned,and the intra-class related dictionary and the difference structure dictionary are obtained. Secondly,when occlusion image classification,the original sparse coding does not consider the coding error,and can not accurately describe the coding error caused by occlusion. In this paper,the error caused by occlusion is represented by a single matrix,which can separate from never oc-clude the feature matrix of the training image. An image of the clear image sentiment classification stage can be recovered by subtrac-ting the error matrix from the test sample. The double-dictionary collaborative representation is used to decompose the clear image samples into identity features and expression features,and finally realize the classification according to the contribution of the expres-sion features of each category in the joint sparse representation. The occlusion experiments in the CK+and KDEF expression databases show that the proposed method is robust to the recognition of random occlusion facial expression images.%为了有效克服遮挡对表情识别带来的误差,同时为了减少表情识别对识别个体的依赖性,本文提出一种基于非凸低秩分解双字典误差模型的遮挡表情识别方法.首先,针对核范数近似秩函数不能有效的估计矩阵的秩,利用非凸对数函数来近似秩函数可以提高估计精度以及对噪声的敏感度.通过非凸对数函数低秩分解将每类表情图像中的表情特征和身份特征分离开,对两部分特征进行字典学习,得到类内相关字典以及差异结构字典.其次,针对遮挡图像分类时,原稀疏编码没有考虑编码误差,无法准确地描述遮挡带来的编码误差,定义单个矩阵来表示由遮挡引起的误差,该矩阵可以从未遮挡训练图像的特征矩阵中分离出来.通过从测试样本中减去误差矩阵可以恢复出清晰的情感分类阶段的图像.利用双字典协同表示将清晰的图像样本分解为身份特征和表情特征,最终根据各类别表情特征在联合稀疏表示中的贡献量进行分类.在CK+和KDEF表情数据库的遮挡实验结果表明,这种方法对随机遮挡表情图像的识别具有鲁棒性.