Sparse representation guided low-rank restoration for noisy image recognition

Abstract

Image recognition is a crucial task in computer vision. However, the recognition accuracy is highly depended on the quality of the image. The images we acquired are often inevitably corrupted. For noisy images, traditional methods often first restore the image by denoising, and then send the restored image to the classifier. However, this straightforward approach treats recovery and recognition as two independent processes, resulting in a lack of information interaction between each other. In this paper, we propose a sparse representation-based joint blind image denoising and recognition method to address the difficult problem of identifying noisy images. The algorithm iterates between the denoising and recognition processes to find a denoising state that can be identified for the purpose of noise image recognition. Based on this sparse low-rank prior, we demonstrate that information interaction for image denoising and recognition tasks can benefit each other. Extensive experiments on two face databases with different noises demonstrate that our algorithm is more efficient and robust than traditional methods that deal with these two tasks independently.