Robust Neighborhood Preserving Low-Rank Sparse CNN Features for Classification

Convolutional Neural Networks (CNN) has achieved great success in the area of image recognition, but it usually needs sufficient training data. Meanwhile, similar images tend to deliver compact CNN features, so the original CNN features of different images of each subject or similar subjects should have the low-rank and sparse characteristics. Moreover, CNN features may contain redundant information and noise. To this end, we investigate how to discover the robust low-rank and sparse CNN features and show how these features behave for image classification, specifically for the case that the number of training data is relatively small. Specifically, we perform the robust neighborhood preserving low-rank and sparse recovery step over the original CNN features so that salient key information can be extracted and the included noise can also be removed. To demonstrate the effectiveness of the computed joint low-rank and sparse CNN features on image classification, three deep networks, i.e., VGG, Resnet and Alexnet, are evaluated. The simulation results on two widely-used image databases (CIFAR-10 and SVHN) show that the extracted joint low-rank and sparse CNN features can indeed obtain the enhanced results, compared with the original CNN features.