New Method For Digital Image Inpainting And Its Application

As a type of information symbol, images have great superiority over texts and becomemore and more important mode of information expression. However, data missing phenomenaare very common in our daily lives and science research, such as damaged images andocclusion. How to repair the damaged regions becomes an attention. At present, there are twothemes closely related to this topic, one is image structure restoration, and the other is texturesynthesis. How to restore the missing information on natural images is the focus of thisdissertation. In order to obtain this purpose, we carry out some research on the principle,methods and applications about image structure restoration and texture synthesis.We first introduce the background, importance and necessary of our research and indicatethe purpose and content of our research. And then we introduce the concepts, evaluation andbasic idea of the image inpainting and its development conditions. Then we analyze the modelsand advantage/disadvantage of typical image inpainting methods including: image inpaintingbased on partial differential equations (PDEs), image inpainting based on texture synthesis,image inpainting based on improved block mapping and image inpainting using radial basesfunction (RBF). After these, we propose structure and texture image inpainting based on regionsegmentation. The presented algorithm including several steps: in the first step, we decomposethe original image into structure and texture components. In the second step, we connect thebroken edge based on tensor voting and then segment them into independent regions. In thethird step, we restore the structure and texture separately in every region. Finally, we add therestored structure and texture images together and get the inpainting result.The research results of this paper prove that the image structure and texture decompositionand segmentation of the broken region may be a new direction for digital image inpainting. Thestructure and texture components of image are independent, which is good for researchingspecial algorithm for each of them. And different objects belong to different regions aftersegmentation. This will not introduce unwanted information, which can see from theexperiment results that our method is more robust and more suitable for natural image.