| Image inpainting refers to restore the lost or damaged area according to the information of the known area in a image. In recent years, with the widely application of image inpainting in varieties of problems, such as restoring images from text overlays, object removal in a context of editing, video loss concealment and so on, this technology has become the attention points and research hotspot of the domestic and foreign image workers.In mathematics, image inpainting is considered as an ill-posed problem that has no well-defined unique solution because of the lack of adequate image prior information.In order to solve this problem, it is necessary to introduce varieties of image priors.Image inpainting methods can be divided into diffusion based method and texture synthesis based method according to the difference of the image priors we assume. The diffusion based inpainting focus on the smoothness priors via establish and solve partial differential equations(PDEs) to propagate local structures, in which the most studied is the BSCB model and TV model. The texture synthesis based inpainting considers the self-similarity priors via search and replicate the best matching sample patches in the known area to complete the repair, in which the exemplar-based model has obvious advantages in the repair efficiency and effect. This paper makes simulation experiments for all the inpainting modes mentioned above base on the analysis of their principle, and then proposes some improvement and innovation of the existing problems in the classical model, mainly:(1) This paper improved the diffusion factor of the p-Laplace model to obtain an adaptive variational inpainting model. The model is proved to be able to switch between the TV model and the thermal diffusion model with the adaptive regulation of the size of the diffusion factor by analysis its diffusion characteristics. In the flat area of the image, the model is similar to the thermal diffusion model, which can effectively avoid the staircase effect in TV model. In the edge area of the image, the model is similar to the TV model, which can effectively keep sharp edge of the image. The simulation results indicate that the proposed model in this paper has improved the convergence speed and PSNR compared with the TV model.(2) This paper points out some deficiencies of exemplar-based model, and putsforward some improvement methods: To ensure the robust of the the calculation formula of priority, using the weighted sum of the confidence item and data item instead of product. To ensure the accuracy of the best matching sample patches, using structural similarity to replace the sum of squared differences as the similarity metrics. To ensure the visual connectivity of restoration image, using smoothing filter on the damaged region boundary to eliminate the effect of the joint after the sample patches have been replicated. The simulation results indicate that, the improved model proposed in this paper performs better in propagating the image structure and more natural in the vision.All of the above is the research work we have done around image restoration in this paper, image inpainting is closely linked with our life, worthy of our further study. |
