Hash Based Multi-source Image Patches Retrieval And Its Application

The patch based image modeling has been widely used in low level vision problems.A series of image processing methods using nonlocal self-similarity prior need to search for high quality similar patches.Focusing on denoising,we find that most existing methods use patches intercepted from degraded images with noise only,ignoring the large quantity of clean and similar patches in external images.As a result,a hash based multi-source image patch retrieval method is proposed in this paper.With the combination of the nonlocal self-similarity prior,the hierarchical sub-inclusion relationship of patches and the coherency hypothesis of smooth region of images,we divide the problem into two parts.On one hand,the proposed method is an image retrieval problem,using appearance similarity as metric in image level.On the other hand,it is a dense patch matching problem between similar images in patch level.In this paper,the appearance similarity is defined as the quantity of similar patches between two images.We train an end-to-end deep hashing network with pairwise labels using texture images with high appearance similarity.As proposed,the network we train is able to learn appearance similarity among images.Moreover,the image containing the patch to be retrieved can be quickly located after being encoded by the hashing network.With the improved coherency sensitive hashing,similar patches are propagated in smooth region.In non-smooth region,they are mapped into the same hash bins by Local Sensitive Hashing.Finally,boosted by the nonlocal self-similarity prior,more similar patches can be selected with high precision.Denoising is performed in the simplified BM3 D framework by replacing patches from the noisy image with the collected patch groups.The testing results show comparative PSNR and good visual quality compared with many state-ofthe-art methods.