Research On Learning-Based Low-Level Vision Problem

Recently, computer vision as the hot research topic has received more and more attention from the computer scientists, especially the fundamental researches relating to low-level vision, such as image super-resolution, image restoration, and noise alleviation. Also low-level vision has vast demand of real world application, ranging from astronomy, remote sensing image, and medical image to network applications. Therefore, it has become an active research topic with great application requirement.In this paper, we mainly discuss the two traditional problems: image super-resolution and image restoration for compressed image. Both of these two techniques are targeting providing higher resolution and higher quality image to the observers. Consequently, it is the problem relating to the low-level vision, in which analysis and understanding of the image content is not relevant and considered. These days, machine learning approaches have been proposed and provided some thoughts for the traditional computer vision problems. Thereby, learning-based approaches for solving computer vision problem have been researched by the scientists. In this thesis, learning-based approaches for low-level vision are to be discussed.(1) Learning-based image hallucinationWe analyze the key technologies of traditional image super-resolution and each exiting problem. The traditional learning-based image hallucination always introduces irregularities into the hallucinated image. Therefore, we propose a three tiered network model to remove the irregularities. Firstly, the hallucination with primal sketch priors is performed to construct a coarse high-frequency component. Secondly, enhancement is implemented to enforce local compatibility between the patches in the constructed component. Thirdly, a Markov network is utilized to refine the enhanced high frequency component. Experiments demonstrate that our model can hallucinate higher-quality images than existing methods.(2) Learning-based image restoration for compressed imageIn this paper, we propose a novel learning-based image restoration scheme for compressed images by suppressing artifacts and recovering high frequency components with the priors learned from a training set of natural images. Specifically, deblocking is performed to alleviate the blocking artifacts. Moreover, consistency of the primitives is enhanced by estimating the high frequency components, which are simply truncated during quantization. Furthermore, with the assumption that small image patches in the enhanced and real high frequency images form manifolds with similar local geometry in the corresponding image feature spaces, a neighboring embedding-based mapping strategy is utilized to reconstruct the target high frequency components. And experimental results have demonstrated that the proposed scheme can reproduce higher-quality images in terms of visual quality and PSNR, especially the regions relating to the contours.