Study On The Digital Image Inpainting Algorithms

Digital image inpainting technology research is the currently popular direction in the computer graphics and the computer vision. Degraded images are inpainted by the prior knowledge of degradation, that is to say, the inverse degradation process of digital images. Its research region is not only limitted to inpaint degraded images, but also used to remove the object or text in an image for a specific purpose.In this paper, this research is deployed on three points as follows.Firstly, the basic theory of digital image inpainting technology is introduced, such us the basic principles of image inpainting, the best guess principle, the Bayesian framework theory, the evaluation standards of image inpainting, the variational method and the functional extremum, etc. Then, three image inpainting models based on the Partial Differential Equation are studied, including the Bertalmfo-Sapiro-Caselles-Bellester inpainting model, the Total Variation inpainting model and the Curvature-Driven Diffusions inpainting model. Experimental results show that the performance of three models mentioned above is not well as expected in rich texture areas because of just the structure information of digital images be used.And then, a problem is that the performance of the Total Variation inpainting model and the Curvature-Driven Diffusions inpainting model is closely relevant with the complexity of inpainting images and inpainting areas. According to the analysis above, a new image multi-level-inpainting method is proposed for improving the inpainting effects, including two aspects'research contents. On the one hand, the Total Variation model needs be improved in the vision connectivity. Therefore, the Curvature-Driven Diffusions model is used to combine with the Total Variation non-denoising model to deal with the gray image. On the other hand, in order to improve the inpainting effects of the Color Total Variation model in rich texture areas, the Curvature-Driven Diffusions model is combined with the Color Total Variation denoising model to deal with the color image. Experimental results show that the new image multi-level-inpainting method performs better both in the structure and the texture. Moreover, the method not only can be used in the image inpainting with largerer damaged areas, but also can be used to inpaint images with an undetectable manner. This new image multi-level-inpainting method is robust to different inpainting images and inpainting areas.Finally, image inpainting algorithms based on texture synthesis scheme are studied, pointing two problems existed in the paper of the region filling and object removal by exemplar-based inpainting proposed by CRIMINISI A.et al. Specificlly speaking, when calculating the priority of pixels, the different boundary pixels should be used in different sampling areas, and the different size of matching templates also should be used. On the basis of analysis, a new adaptive selecting method applied to the texture synthesis is proposed. In this paper, an adaptive gradient factor is introduced to improve the choice of matching the size of the texture blocks. Experimental results prove that the adaptive selecting method not only performs better in texture detail at the same time of maintaining the structure information, but also better in the speed of image inpainting.