Superpixel-guided nonlocal means for image denoising and super-resolution

Abstract

The dramatic growth of online multimedia data has resulted in a great demand for high-quality images. The two most required multimedia content analysis applications are image denoising and image super-resolution. The effective nonlocal means (NLM), which exploits the redundancy of small patches in natural images, has been applied to image denoising and super-resolution (SR). However, a square window used in the NLM weight estimation may be ill-suited for edge regions, besides which the window size often requires an empirical study to be conducted on test images and is fixed for all the pixels. To adaptively select the neighbors with higher matching precision, we propose a novel superpixel-guided nonlocal means (SNLM) algorithm. Utilizing the superpixel segmentation method, we divide an input image into many small regions, each of which has similar luminance and color values and adjacent positions. One or more superpixels are chosen as the search region for each pixel. In this paper, similar local patches can be found in the selected superpixels rather than using a square window, and can then be used for the weight estimation. The thorough quantitative and qualitative results demonstrate that SNLM is more effective for image denoising and super-resolution than the conventional NLM-based method.