The Research On Image Inpainting And Video Inpainting Based On Texture Synthesis

Digital image and video inpainting mainly aims at filling in missing pixels in an unknown region of an image or video in a visually plausible way. And it is also one of the important research fields in computer graphics, computer vision and image processing. Currently, the primary methods of image inpainting include "Texture Synthesis" algorithms and the algorithms based on Partial Differential Equations (PDEs). On the basis of texture synthesis, this thesis carries out some researches on image inpainting and video inpainting respectively.The exemplar-based inpainting method searching the source patch globally easily leads to some matching errors, and such errors will accumulate and propagate into other areas quickly, even making the results completely unacceptable. Focusing on this disadvantage, the thesis proposes an algorithm for digital inpainting based on color region segmentation. The presented algorithm based on MRF model uses the K-mean clustering algorithm to segment the damaged image, and limits the patch matching within the relevant regions which are overlapped by the current patch. Experimental results show that this method can efficiently handle different cases, especially the large regions in complex background. Furthermore, this thesis extends the image inpainting method to video inpainting. In general, the novel algorithm is applicable to restoring damaged video sequence whose background is immovable and the foreground approximates to periodic motions, and achieves reasonable restoration effects.Image quality assessment plays an important role in analyzing inpainting algorithm performance. At present, PSNR is the mostly used objective evaluation index, but it doesn't often correlate to how the human visual system detects and responds to image inaccuracies. The thesis proposes a new index based on SSIM, in which several factors that can cause decrease on image quality are considered. These factors include image luminance, clarity and gradient magnitude. The experimental result indicates that the index is in accordance with human subjective visual sense.