Research For Image Super Resolution Reconstruction Based On Dictionary Learning

Image super-resolution reconstruction is a technology to improve image quality by using the software algorithm,which can recover the degrade image to the real appearance of the original image as well as possible. It reduces the cost of high-resolution images obtained by hardware greatly, and plays an important role in improving the visual perception. At present, super-resolution has been widely applied into computer vision, medical imaging processing, security monitoring and military image processing. super resolution reconstruction methods based on dictionary-learning are very popular in recent years. These methods have achieved good results, but there are still some problems need to be solved for example, the high time-consuming of dictionary training process, the limit to preserve the edges, visual artifacts and non-full use of the prior knowledge of the LR image itself. In order to overcome the weaknesses of the super-resolution reconstruction based on dictionary learning, this thesis carried out further research work based on dictionary learning in the super-resolution; Thesis’s main work and innovations are summarized as follows:1. The method of image super-resolution reconstruction via Improved Dictionary Learning Based on Coupled Feature Space is studied. In the proposed algorithm, at first, the Gaussian mixture model clustering algorithm is employed to cluster the training image blocks, secondly, quickly obtain high and low resolution feature space of dictionary and mapping matrix by Boost KSVD dictionary learning algorithm, and then, the Super-Resolution image is reconstructed according to the likelihood probability of test samples, in which each category adaptively selected the most matching dictionary and mapping matrix for high-resolution reconstruction, finally, the non-local similarity and iterative back-projection are exploited to furtherly improve the quality of the reconstruction image. The experimental results demonstrate the validity of the proposed algorithm.2. In order to overcome the weak of the limit ability of preservation of edges and easy to produce visual artifacts in some super-resolution methods based on dictionary learning, the method of multi-dictionary learning image super-resolution method with edge-enhanced is studied, the method can effectively restore the image edge detail. Firstly, the training image patches will be classified by using K-means, and be quickly learned multi-dictionary pairs by employing the Boost K-SVD algorithm, during the super-resolution reconstruction, the method of this paper adaptively select the optimal dictionary pairs for sparse decomposition and recovery. To improve the visual quality of edge after image reconstruction, we employed direction-preserving regularization according to the input test low-resolution image, meanwhile learning the natural image database edge sharpness statistics prior to constraint the image reconstruction of edges. Through experiments for pictures demonstrate that the proposed algorithm has remarkable improvement in peak signal-to-noise ratio, structural similarity and visual quality compared with the other learning-based algorithms.