Sparse Representation Based On Single Image Super-resolution Algorithm And Dictionary Learning

With the rapid development of video technology and network communication technology, the improvement of image quality and increased speed of transmission image have become an urgent demand of the people. Super-resolution technology isn’t need to increase the number of hardware, but uses image processing methods to reconstruct low-resohrtiorr images into a high-resolution image, increasing the high-frequency details of the image. However, when we need a larger magnification or we only obtain a small number of low-resolution images in the same scene, the image will be restored fast degradation by the traditional interpolation methods and the reconstruction methods.To solve this problem, based on the theory of compressed sensing, namely in mild conditions, the sparse representation can be accurately restored from the down-sampling signals, this thesis reconstructed a high resolution image from a single frame low resolution image with the method of sparse representation and dictionary learning. Image statistical studies show that the image blocks can be well represented as a sparse linear combination of some elements, which are chosen from a suitable overcomplete dictionary. Based on this finding, this thesis firstly obtained a high resolution dictionary and a low resolution dictionary by dictionary learning, then find a sparse representation coefficient of the input low resolution image blocks. Secondly, we can reconstruct a high resolution image with the use of the high resolution dictionary and the solved sparse representation coefficients. The main works are as follows:(1) This thesis studied super-resolution algorithms based on sparse representation and dictionary learning. In my works, the high and low resolution dictionaries were trained by K-SVD method, and coefficients of the sparse representation were calculated by OMP method. Then, a high resolution image was reconstructed with the use of a high resolution dictionary and the solved coefficients of the sparse representation.(2) In order to improve the accuracy of super-resolution images, this thesis used the joint dictionary training methods to strengthen the isomorphism between the high and low resolution dictionary, guaranteeing the sparsity priori constraints. During the dictionary training, this thesis extracted gradient features of low resolution image blocks to improve the sparse representation capacity of elements in the dictionary. Finally, the thesis used a global reconstruction model to optimize the reconstructed image, which can eliminate local differences.(3) In order to improve the speed of super-resolution image restoration, this thesis proposed a fast sparse representation algorithm based on image blocks selection strategy. Based on the Variance threshold value, the thesis used the sparse representation algorithm to deal with the blocks at the edges and contours of the images, and used bicubic interpolation algorithm to deal with the blocks in the smooth regions, which can balance the image quality and the speed of restoration.(4) The simulation experimented in the Matlab platform, experimental results proved that the ideas and algorithms presented in this thesis were effective. Compared to the bicubic interpolation algorithm, the sparse representation algorithm proposed in this thesis had great advantages in image accuracy., and was still better than the similar sparse representation algorithm. Finally, the thesis discussed the robustness to the noise of the sparse representation algorithm, and analyzed the relationship between the size of the dictionary and the image quality.