Impainting with a Nonlocal Means Filter

Abstract

The need to fill areas of an image distorted by artifacts with texture from undistorted areas is called impainting. Impainting is used both to improve the visual perception of an image and in classical recognition and robotics problems in order to remove irrelevant information from an image. Modern methods of impainting use neural networks. However, these approaches have drawbacks that do not allow the use of these algorithms in the practice of computer vision. In this article, we propose to use a nonlocal means (NLM) filter, which has proven itself to be excellent in image noise reduction tasks. The basis of our motivation is the fact that the goal of the NLM, like any method for recovering images distorted by noise, is to minimize the distance (or error by some criterion) between the original image and the reconstructed one. The result of computer experiments showed that the proposed method of impainting is superior to other methods according to peak signal-to-noise ratio (PSNR) criterion. The effectiveness of the proposed filter is also shown with the help of illustrations to the article, so that the reader can compare the quality of different processing options visually.