基于块稀疏递推残差分析的稀疏表示遮挡鲁棒识别算法研究

摘要

A robust occlusion pattern recognition algorithm is proposed which considers how to detect occluded region automatically with unknown occlusion pattern and conquers the influence of occluded region to improve the robustness of the recognition algorithm. Firstly, the test image is divided into up module and down module. Next, the sparse representations are computed respectively. Then, the module with higher sparsity and the corresponding sparse solution are found. The test image is reconstructed using this module and the N largest coefficients. According to the residual of original test image and reconstruction image, the occluded pixels can be confirmed coarsely. Considering the continuity of occluded region, the coarsely confirmed occluded region is regularized by morphological operation and gets the weighting matrix. Finally, the test and training set are weighted and normalized by using this weighting matrix and then the final decision is made by using global sparse representation. The experimental results on AR, Yale B and MNIST databases verify that the proposed method can detect the occluded region roughly, and the effectiveness and the robustness of the proposed method can be observed obviously.%针对如何在未知类别的情况下自动检测出遮挡区域，然后在克服遮挡影响的基础上提高识别算法的鲁棒性问题，提出基于块稀疏递推残差分析的稀疏表示遮挡鲁棒识别算法。该算法首先将待测样本分为上下两部分，并分别用对应块的训练样本进行稀疏表示，找出稀疏度更高的块及对应的稀疏解，并将更稀疏前N个解推广到另一个块中，重构测试样本。然后根据重构测试样本与原测试样本的残差推测遮挡像素。考虑到遮挡区域的连续性，利用形态学操作对推测的遮挡区域进行规则化处理并得到加权矩阵。最后利用加权矩阵对测试样本和训练样本进行整体加权归一化，再利用全局稀疏表示进行最终的分类判决。在AR、Yale B及MNIST上的遮挡仿真实验证明该方法不但可大致确定遮挡区域，还可提高遮挡图像识别的性能。