A Parallel Algorithm for Multi-view Image Denoising

Abstract

In this paper, we propose to improve the denoising performance by exploiting the redundancy provided by the multiple views. Here, we review the rich literature on image denoising methods. Zhang et. al. in [1] conjecture that single view image denoising algorithms have reached the limit of their performance. [1] also formulates the problem of multiple view denoising and give an algorithm to achieve the same. Their results prove the improved performance attained because of the information in the additional views. In this paper, we propose a novel approach towards this problem of denoising images using multiple views. We use an adaptation of the NL-means denoising algorithm on images focused at different depths or as we call them, focal images. These focal images are constructed using the multiple views. We introduce the notion of super pixels that constitute the focal images. The NL-means denoising algorithm denoises these super-images. Depth values are simultaneously estimated. Each view is reconstructed using these denoised super-images and the depth map. This intuitively parallel algorithm is implemented on GPU. We present the details of our implementation. The results of our experiment not only validate our hypothesis of improved performance due to multiple views, but also show that our GPU implementation is faster than other algorithms which have comparable performance. We compare the performance of our algorithm with the state-of-the-art single view image denoising and multiple view denoising algorithms.