Multilayer Image Inpainting Approach Based on Neural Networks

This paper describes an image inpainting approach based on the self-organizing map for dividing an image into several layers, assigning each damaged pixel to one layer, and then restoring these damaged pixels by the information of their respective layer. These inpainted layers are then fused together to provide the final inpainting results. This approach takes advantage of the neural network's ability of imitating human's brain to separate objects of an image into different layers for inpainting. The approach is promising as clearly demonstrated by the results in this paper.

I. Introduction

Deteriorations of images and pictures are often caused by aging effects, bad maintenance and other wearing or tearing problems. Partly damaged pictures can be inpainted by human painters. With the rapid development of modem processing and computer-vision technology, it is now feasible to develop digital techniques for performing the same task. Many researchers have attempted to find suitable methods and algorithms to efficiently restore images from their partly damaged versions by computers automatically [1]. Automatic digital inpainting is a technique which restores damaged image or video by means of image interpolation [2]. The technique not only can inpaint damaged pictures, but can also be used in removing a user-selected region or object [2]. Image zooming can also take advantage of this technique.