Small Scale Damaged Image Inpainting And Its Application

Due to the development of digital image technology in recent years, digital image inpainting technology has become an important research field of image processing. According to the field of image inpainting application, it can be divided into two parts:image inpainting algorithm for small damaged scale and filling algorithm for large area.This thesis selects image inpainting algorithm based on small damaged scale to study, including image inpainting algorithm based on Normalized Convolution(NC) and total variation(TV). On the basis of studying the existing normalized convolution image inpainting algorithm, the sawtooth effect and the superposition error have been found. The reason of sawtooth effect is only using spatial information when calculating the neighborhood weight matrix, but not considering the feature information of image. In this thesis, the neighborhood weight matrix of a damaged pixel to be estimated is constructed by computing the isophote information, to ensure to repair the damaged area along the isophotes direction. The cause of superposition error is that the image restoration algorithm based on interpolation is a greedy algorithm. In the algorithm the pixel which be inpainted prior has large influence on the pixels which be inpainted after. The window radius of damaged pixel to be estimated is adaptively selected according to the distance between the damaged pixel and the damaged boundary, and the proportion of image known information is increased, to reduce the superposition error. Experimental results show that the proposed algorithm can better maintain boundary and texture information.Image inpainting algorithm based on total variation(TV) is a diffusion-based algorithm. The basic idea is to diffusion the neighborhood information of damaged area to the damaged area. This algorithm is sensitive to the selection of parameters, and the convergence speed is slow. This thesis proposes an adaptive image inpainting algorithm based on total variation. The value of paramater is selected adaptively by the neighborhood information of pixel which to be estimated. In the quality of inpainting condition, this algorithm has the faster convergence speed than total variation. At the same time, in the study of TV model algorithm, smooth effect was found. The smoothing effect of TV model algorithm is used to modify the result of normalized convolution, so as to ensure to smooth the boundary and have a good visual effect.Finally, this thesis applies the results of image inpainting to the self-embedding watermarking. In the self-embedding watermark by using the different block size and tamper ratio, the blocks of synchronous tampering were inpainted. Experimental analysis shows that image inpainting technology can effectively solve the tamper recovery of self-embedding watermarking.