Research Of Image Inpainting Algorithm Based On Exemplar And Low-rank Theory

With the rapid development of computer vision, the requirement of the image quality is getting higher and higher. When the image is damaged and cannot provide enough information, digital image inpainting technology emerges. Image inpainting is a computer technology which could make the damaged image recover as itself. Its principle is using the known pixels to estimate the unknown pixels in the image. Its objective is to repair the damaged region or remove the specified object. The repaired image needs to satisfy the visual requirements of observers and make the observers cannot detect the mark of inpainting. At present, image inpainting technology has been applied to the recovery of arts, photography, traffic surveillance, video production and so on. It has theoretical significance and practical value.This paper studies some classical algorithms in the digital image inpainting technology. They are diffusion-based, exemplar-based and lowrank-based algorithm. And we indicate their advantages and disadvantages. Among these algorithms, the exemplar-based inpainting is the most deeply researched and its development is quite rapid. The Criminisi algorithm repairs the missing region at the patch level. Its computation of priority is not precise enough; mismatching always appears in the patch searching. The repaired result is not ideal.In order to meet the demand of human vision and improve the quality of the repaired image, this paper proposed a new image inpainting algorithm based on the framework of exemplar-based inpainting. The new algorithm obtains the structure information according to the structure tensor during priority computing; two-step searching method could decrease mismatching as much as possible; the unknown pixels are estimated by the strategy which combines similar structure and low-rank theory. The experimental results show that the proposed algorithm can set reasonable patch priority, reduce the probability of mismatching, and maintain the completeness of structure. The repaired images have more natural appearance than the results of some other mainstream image inpainting approaches.There is a problem in the existing inpainting algorithms. The damaged region needs to be marked by hand which is not convenient to the algorithms. In the text removal, this paper proposed an effective English detection method which could locate the text accurately. Compared with the artificial marking, the automatic detection is faster and the location is more accurate. The new detection algorithm not only reduces the complexity of text removal, but also improves the visual effect of the results. At last, we apply our image inpainting algorithm to the video inpainting by proper transforming. The results of video inpainting are satisfactory.