The Research On Digital Image Inpainting Technology

Digital image inpainting technology use the known information in the damaged image to fillthe defect area. It belongs to the image restoration problem in the field of computer vision.Image inpainting has always been one of the key research direction in computer vision, and it istypically used for computer multimedia editing、removal of person or text and the protection ofthe privacy information in image. After years of development, this technology has beengradually mature. Although more and more inpainting algorithms have been proposed, perfectingthe theory of image inpainting and the pursuit of better inpainting results have always been theideals and goals for research workers.First of all, this paper introduces the research background and significance of digital imagerestoration technology. Then, according to the optimal type of image inpainting algorithms, wedivide existing inpainting algorithms into two categories, that is: the inpainting algorithms basedon local optimization and global optimization.The focus of our research is the image inpainting algorithms for large defect areas, in orderto highlight our research work, the second chapter firstly introduces the inpainting algorithm forsmall defect areas, that is the image inpainting algorithms based on partial differential equation.We introduce the three classical inpainting models in partial differential equation respectively:BSCB、TV and CDD. However, restricted by theoretical framework, the methods based onpartial differential equation are mainly used for inpainting the damaged areas of small scale. Sothe third chapter starts to introduce Criminisi algorithm. It performs well in inpainting thedamaged images of large scale. However, for those images with strong structure, Criminisialgorithm obtain poor visual connectivity in the result. Against the shortcomings, this paper putsforward a improved algorithm. First of all, we design a new priority function, which strengthensthe ratio of structure component in the defect area, in order to strengthen the structure continuityin the inpainting results. Secondly, we adopts a faster searching method with morecomprehensive patch matching criterion, in order to obtain the more reliable matching block andspeed up the inpainting rate. Finally, we verify the effectiveness of our improved algorithm through experiments.Although Criminisi algorithm can receive better inpainting effects compared with themethods based on partial differential equation for large defect areas. In essence, these twomethods are all local optimal inpainting algorithms. Sometimes, for specific damaged images,the overall visual effect in the inpainting results is poor. The fourth chapter in this paperdiscusses the Komodakis inpainting algorithm based on global optimal. This algorithm is basedon the MRF model, and uses Priority-BP algorithm to optimize its global energy function. Itsinpainting results are more consistent with human visual characteristics. After analyzing theinpainting principle and steps of Komodakis algorithm, aiming at the threshold problem for labelscreening in this algorithm, we propose the adaptive threshold label screening inpaintingalgorithm based on the MRF model. First, We use MSD (Mean Square Deviation) to assess thematching degree for each MRF node’s label set, then we can set the corresponding thresholdsadaptively. Through this we can get more reasonable priorities of nodes and more accuratemessage scheduling. Second, we also evaluate the size of label set for each MRF node underdifferent clusters combined with Kmean algorithm, then we can make dynamic label cutting foreach node adaptively, through this we can enable the iteration to work more effectively. Finally,we verify the effectiveness of our improved Komodakis algorithm through experiments.