Research Of Image Inpainting Based On SOM Neural Network

Image inpainting is very popular with scholars for its use value and broad application. As many of the electronic wares are no longer too expensive to people, needs on image inpainting is rising fast. Image inpainting belongs to image processing. It mainly consists of computing the lost information of an inpainted image by ambient information. With the development of computers and efforts of many researchers, many excellent image inpainting algorithms have been such as PDE (partial differential equations) based inpainting method and texture-based completion method. These two algorithms have different application. PDE based inpainting method can deal with inpainted image with small nick while texture-based can give a good result when the damaged area is relatively big. However, they all have their limitationsArtificial Neural Network (ANN) is another hot spot which is studied by more and more people. It has already been applied in many areas, and the result shows that ANN still has potential applications. ANN is also applied in image processing, but not broadly. Back propagation (BP) network, Radial Basis Function (RBF) network, Self-organizing Map (SOM) network are typical representations of ANN.Self-organizing Map is an important branch of ANN. SOM is well known as it can map inputs'topology and implement auto classification after unsupervised learning. In this paper, SOM is employed to simulate the relevant structure lines. Then SOM is also employed to classify the relevant structure lines while there are two structure lines in the damaged area.This paper reviews the existing image inpainting algorithms and proposes a new image inpainting method based on SOM neural networks, and improve the proposed method. The main contributions of this paper are as follows:1. This paper presents a structure information preference image inpainting method based on self-organizing map (SOM), and the efficiency of the proposed method is demonstrated by the simulation results.2. The proposed method is improved and more complicated damaged image in which there are two structure lines in the damaged area is studied and perfect simulation results are provided by case studies.