Sparse representation has been deeply studied for its robustness and effectiveness,while its complex computa-tion reduces the efficiency;so many methods combined with nearest neighbor,subspace learning and sparse representation to reduce computation are proposed.Considering the intra-class scatter is small while the inter-class scatter is large in the projection subspace,and samples of a same class have similar contributions to reconstruct a given sample,so dealing with the problem by class is more reasonable.Weighted nearest neighbor classes based block-sparse representation for image recognition is presented in this paper.First,k neighbor classes are selected and the corresponding coefficients are pre-served to weight samples of the k classes,and then the test sample is represented by block-sparse representation for classifi-cation in the subspace.Finally,it classifies the test sample into the class with lowest residual.The experiments on AR, Yale B,MNIST and PIE database verify the proposed method’s effectiveness and robustness.%针对如何将近邻、子空间学习与稀疏表示结合起来解决基于稀疏表示的图像识别问题，本文综合考虑子空间中样本的类内散度小，类间散度大，且同类中所有样本对重构某一给定样本的影响相似（即表示系数相似），因此按类而非样本处理的思想更符合基于类重构误差进行分类的算法要求，为此提出一种基于近邻类加权结构稀疏表示算法用于图像识别。该算法首先利用线性类重构误差选取k个最近邻类，并将其对应的系数作为权值对投影后的近邻类加权，其次在投影子空间上，用k个类的加权训练样本集对测试样本进行结构稀疏表示，最后根据最小类重构误差得出分类结果。在AR，Yale B，MNIST，PIE数据库上的实验结果表明该方法在训练样本数较少的情况下获得较高的识别率且具有一定的鲁棒性。
展开▼
