Application Of Sparse Representation In Image Quality's Improvement

In the course of signal formation, transmission and acquisition, the image will inevitably produce degradation in the quality affected by the hardware devices and transmission channel. Image restoration is to recover the high quality image with the utilization of the image's prior knowledge, and is widely used in remote sensing images, medical image analysis, video monitoring and others. The main contents of this dissertation can be listed as follows:Image restoration is based on solving an optimization problem with adding some appropriate constraint terms. Studies have shown that the constraint term of natural image's prior has a good effect for improving the quality of the image to be recovered. Most of the existing image restoration algorithms fail to take into account the geometric structural features within the image, so the quality is affected. One of the important priors is the local image structures. Steering kernel regression is adapted locally to image features. Thus, we use it to characterize the image's local self-similarity. The local self-similarity just analyzes the pixels in the neighbor areas; however, pixels far away from each other also have similarities. Considering the non-local redundancies, we used non-local self-similarity as another constraint term. In the process of image degradation, the most lost information is the edge structures. If we are able to recover the edge information, high quality reconstruction image can be obtained. Therefore, we established a new reconstruction model with local self-similarity, non-local self-similarity and edge structures as constraint terms, and reconstructed high quality images.We proposed a super resolution reconstruction algorithm based on high frequency (HF) and middle frequency (MF) information. The main ideal is to reconstruct the lost high-frequency according to the mid-frequency of the image. This algorithm selects image's HF and MF as patch pairs, which are trained by generalized iterated shrinkage algorithm to obtain a HF and MF dictionary pair. According to the MF of test images and the dictionary pair, image HF are reconstructed. Then, combined with interpolated low resolution (LR) images and reconstructed HF, high resolution (HR) images are reconstructed using a non-local similarity regularization term. Our algorithm is helpful to learn the relationship between the LR image and the HR image. Experimental results show that the proposed method has a sharper edge.The dictionary learning plays an important role in reconstructing high images. However, many existing algorithms of dictionary learning also have some restrictions. Nonparametric Bayesian method can build a model towards observed data which meets certain distribution, and it has been widely used in many areas. So we introduced Bayesian nonparametric sparse representation model of Dirichlet-Multinomial process and Beta-Bernoulli process into the image super resolution reconstruction and image denoising. In the image super resolution, we combined the above two model with the HF and MF information respectively, and learned two nonparametric dictionaries to recover HF information which is lost. The experimental results validate that our algorithm has the ability to reconstruct high quality images and to reduce computation time.