局部球面规范化嵌入:PCANet的一种改进方案

摘要

When there is high-level illumination change or occlusion in a facial image,the feature maps produced in PCANet will mainly concentrate on the vicinity of 0 and thus lose redundancy in some extent.In order to enhance the ability of PCANet against nois,a local sphere normalization(LSN) method was proposed and embedded into the convolutional layers of PCANet.LSN embedding helps improves the redundancy of PCANet feature maps.The experiments on UMBDB and AR show that the improved PCANet is more redundant and robust,and has better recognition performances than the original PCANet.One important finding is that directly imposing LSN on the input image will lead to unstable recognition performance,while the feature redundancy should be improved steadily by LSN.%当人脸图像中存在较大比例的光照变化或遮挡时,PCANet所采用的局部零均值预处理以及PCA滤波器对噪声的过滤作用将导致所生成的特征图的整体分布主要集中在0附近,在一定程度上丧失了冗余性.为了提升PCANet对抗噪声的能力,提出了局部球面规范化方法,并将其嵌入PCANet的卷积层,从而拓展了PCANet特征的丰富性.在UMBDB和AR库上的实验表明,改进后的PCANet具有更好的冗余性、鲁棒性和识别性能.一个重要的发现在于:特征的冗余性需要在噪声滤除的过程中逐步提升,直接对输入图像施加局部球面规范化可能会导致不稳定的识别性能.