Super-resolution Image Algorithm Based Collaborative Representation

At present, super-resolution reconstruction is an important research direction in image processing, and its application is very extensive. Its purpose is to restore or increase the spatial resolution from the image sequence, thereby improving the quality of the image. In recent years, the super-resolution reconstruction algorithm based on sparse representation is the main research’s directions. Based on sparse representation,this article mainly complete the following work:First of all, considering the difference between structures of image blocks. We design a super-resolution image reconstruction algorithm based on image block classification collaboration representation. Using differences of images’ local features, the image block is divided into two types: detail blocks and contour blocks. Then treated separately:(1)To detail blocks, train dictionaries and reconstruct by dictionaries;(2)To contour blocks, using Bicubic interpolation method, it can be directly reconstructed. Finally, using collaboration represents we can train the cross-mapping semi-coupled dictionary and its corresponding collaboration representation. By comparing the experimental results with the traditional algorithm, it shows that the algorithm is reasonable.Secondly, considering universality of noisy image in the real-life, we design a super-resolution reconstruction and de-noising algorithm based collaboration represents. Based on similar structures of the image blocks and repeating the local structure of the image, we can assume a given low resolution image patches can be extracted a large number of similar blocks from the training database. According to this theory,we can denoise the inputting image. By comparison with other algorithms which based on sparse representation, when the noise is large, the algorithm has advantages.At last, in connection with the complex and nonlinear mappings in the space, we propose a super-resolution reconstruction algorithm with learning nonlinear dictionary based on collaboration representation. According to the flexibility of the image’s spatial structure, and in order to accurately represent the fundamental link between high and low resolution sparse coefficient, we combine with Regression Support Vector Machine nonlinear to train the dictionaries and the hide relationship of structural space. At the same time, to increase the rate it uses Collaboration representation and Image block classification. Compared with other algorithms, the algorithm described in this chapter on the reconstruction results has significantly improved.