Research For Image Super Resolution Reconstruction And Application Based On Sparse Representation

Image super-resolution(SR) reconstruction is the process of generating a high-resolution(HR) image using one or more low-resolution(LR) observations from the same scene. It overcomes the lack of high-resolution images obtained by the hardware costly, which plays an important role in adding the image details and improving the visual perception. Learning-based super-resolution reconstruction methods have been very popular in recent years, most of them have the problem of over-reliance on the training images, and don’t make full use of the prior knowledge of the LR image itself. The problem causes the reconstruction HR images contain some obvious artifacts, and the HR images quality is not high. In order to solve these problems of learning-based super-resolution reconstruction methods, this thesis only uses the LR image itself to promote its resolution by fully exploiting the priori information hidden in the LR image itself. The description of our work is as follows:1) proposed a single image super-resolution reconstruction method based on self-similarity and sparse representation. The method combines sparse K-SVD dictionary learning and nonlocal means, which are used to add the effective information hidden in the same scale and across different scales structural self-similarity into the maximum a posteriori probability estimation framework by two different regularization terms. Then, a local optimal solution is obtained by using the gradient descent algorithm. The experimental results show that our method has a better improvement both visually and quantitatively.2) for the problems in the image super-resolution method based on K-SVD, a new method based on sparse K-SVD is proposed and used for license plate images reconstruction. Firstly, the low-resolution (LR) license plate images and their down-sampling versions are used as samples to train dictionary, which improves the relativity between the dictionary and the LR license plate image. And The sparse K-SVD method is used for dictionary training to obtain a pair of dictionaries. Then, a gradual magnification scheme is used for reconstruction. Experimental results on license plate images show that the PSNR is improved by nearly 0.6 dB compared to the image super-resolution method based on K-SVD, and our method also has a better visual improvement to certain extent.In this thesis, we studied the learning-based super-resolution reconstruction methods, for their shortcomings, we have put forward our own ideas and demonstrate the algorithms we proposed through the experiments. Our algorithms have some contribution to enhance the image quality and improve the visual perception through image SR reconstruction.