Image Inpainting Algorithm Based On Sparsity And Collaborative Representation

Image inpainting is a process of combining the known information of the image and effective algorithm to estimate the missing pixels. It is related to many fundamental image processing problems, such as image compression, enhancement and restoration. It has broad applications in artwork restoration, special effects in film making and the literal of the image. Image inpainting is an ill-posed inverse problem. How to make the most of the prior information is the key to solve the problem. This thesis carries out some exploratory researches on image inpainting.Firstly, a new inpainting framework based on robust CS is proposed, which is a 1l-minimization sparse reconstruction problem. Since the 1l-regularized robust CS obtains favorable effects and the missing pixels among image can be treated as impulse noise,we exactly apply the viewpoint to image inpainting.The results turn out to be better compared with 2l-minimization sparse reconstructionSecondly, in order to take full advantage of the prior contained in image self- similarity, an improved exemplar-based image inpainting algorithm is presented, which searches for the similar patches of the target patch in multiple scales to get more reliable matching patches. The damaged patch is synthesized using the concept of linear weighted aggregation. The inpainting quality proves great improvement due to the using of multi-scale similarity.Last, a new inpainting frame work based on joint dictionary and collaborative representation model was proposed. In this paper we make use of the multi-scale self-similarity of image to enlarge the search area to get more reliable similar patches, which incorporate with a learning dictionary trained in database to form a joint dictionary. In addition, a new inpainting framework based on collaborative representation model was proposed. The experimental results show that this algorithm can realize comparative ideal inpainting effect for different kinds of images, especially for large areas of texture. Moreover, this method achieves considerable improvement in computational efficiency compared to the approach based on sparse representation due to avoiding the complex iteration.