Hierarchical hashing-based multi-source image retrieval method for image denoising

Abstract Image denoising is a fundamental problem in image processing. One prominent image denoising approach is to exploit the non-local self-similarity (NSS) prior of natural images. How to find and utilize similar image patches is the key to the NSS-based method. However, most NSS-based methods only use noisy patches within the noisy image itself, whose number and quality are quite limited. To exploit the similar clean patches in external image datasets, in this paper, we propose a hierarchical hashing-based multi-source image retrieval method. First, based on deep hashing, a multi-source image-level retrieval method is proposed to search clean appearance-alike images from the external dataset. Second, on the basis of NSS prior and image coherence hypothesis, an improved continuous sensitive hashing method is designed from the image patch level to create a dense mapping between the external similar patches and the target patches. Finally, using the searched similar clean image patches, a simplified denoising method is developed based on the 3D transformation and thresholding technique. Experiments on two kinds of image datasets have demonstrated the effectiveness of the proposed method.