Image denoising manages to recover a digital image from its noisy version by exploring the statistical features inside a given noisy image. Most denoising methods perform well at low noise levels but lose efficiency at higher ones. In this paper, we propose a novel image denoising method, which restores an image by exploiting the correlations between the noisy image and the images retrieved from the cloud. Given a noisy image, we first retrieve relevant images based on feature-level similarity. These images are then geometrically aligned to the noisy image to increase global statistical correlation. Using the aligned images as references, we propose recovering the image with patch-level noise removal. For each noisy patch, we first retrieve similar patches from the references and stack these patches (including the noisy one) into a three dimensional (3D) group. We then obtain the noise free (NF) patches by collaborative filtering over the 3D groups. These recovered NF patches are aggregated together, producing the desired NF image. Experimental results demonstrate that our scheme achieves significantly better results compared to state-of-the-art methods in terms of both objective and subjective qualities.

Date Published: 26 September 2013
PDF: 10 pages
Proc. SPIE 8856, Applications of Digital Image Processing XXXVI, 88560A (26 September 2013); doi: 10.1117/12.2022506

Published in SPIE Proceedings Vol. 8856:
Applications of Digital Image Processing XXXVI
Andrew G. Tescher, Editor(s)