Research Of Image Inpainting Algorithm Based On Content-adaptive

Digital Image inpainting is the process of efficiently filling the unknown regions of an image caused by replacement or removal of part of the foreground. It is a crucial technology to static or dynamic image editing. Image inpainting is important in theory and practical application. Its key problem is how to preserve the details of original image, retain the integrity and consistence, and avoid the incomplete fusion of the target object and its surrounding. The effective solutions are decisions of the filling order and the completion strategy, which are the most concerns by many researchers. We present two novel image inpainting algorithms. One is pixel-based, and the other is patch-based. The former is based on non-local means model, the latter is based on multi-scale exemplar saliency.Firstly, Non-local means model-based inpainting algorithm is to deal with linear structures. Its main idea is that the interpolation weight is computed according to the non-local means model. The inpainting order is designated for the edge-directed fast marching method. So the edge structure can be well preserved.Secondly, A multi-scale patch saliency model is proposed with the help of the definition of the saliency. The multi scale strategy introduced can further decrease the saliency between the background and the foreground. The patch priority is computed according to the saliency model. The order of filling linear structure first, then texture is well preserved. We also introduce a random sampling method to search and match the exemplar patch, which can accelerate the filling process. The saliency-model based completion method can process both linear structure and texture image. The experiments have verified the effectiveness of the methods, which can satisfy the application the image editing task.