Research On Super-resolution Reconstruction Algorithm Based On Image Block Multi-level Classification And Sparse Representation

Super-resolution image reconstruction technique provides an effective signal processing method to restore the high resolution image from either a single low resolution image or low resolution image sequence of the same scene based on some a-priori knowledge. Since Super-resolution reconstruction technology can significantly enhance the quality of the image without need in improving the gathering image hardware device, it is of great importance in not only theoretical analysis, but in practical application. Moreover, Super-resolution technology has a wide range of potential application scenarios. The Super-resolution technique from single low resolution image is focused on in this thesis, especially sparse representation based image super-resolution reconstruction. Sparse representation super-resolution is based on theory of compressed sensing data representation model, through the process of establishment of over-completed dictionaries, using local similarity relationship between high and low resolution image patch, to reconstruct of high resolution target images.In this thesis, deep research on sparse representation based super-resolution method is conducted, by comparing the values of the index, research results show that dictionary size and dictionary composition has a great impact on the algorithm's reconstruction quality and reconstruction efficiency. According to the number of dictionary size and dictionary structure, to design an algorithm by changing the dictionary structure can improve the running efficiency of the algorithm as well as ensure the quality of the reconstructed image.According to the local features of the diffident image patches is different and the structural similarity between all kinds of image patches. An improved super-resolution image reconstruction algorithm based on multi-level classified image patches and sparse representation is proposed. In the proposed method, image patches are firstly divided into three different forms by threshold features, then using sparse sub-dictionary show a particular type of image patches, the three forms are treated separately:during the reconstruction process, bicubic interpolation approach is used for image patches of (2N×2N)mn; Image patches of (N×N)min are achieved reconstruction corresponding high and low resolution dictionary; Image patches of (N×N)max are divided into smoothing layer and texture layer for morphological component analysis algorithm, where these parts are both mutually independent, then from each layer corresponding sub-dictionary are achieved reconstruction; finally, the fitting of three kinds of image patches to get the final image. The experiment results show that the proposed algorithm can obtain a more detail in edge patches and irregular structure regions, significantly improves in effect of reconstruction. By comparing the values of PSNR and SSIM, to prove the feasibility and effectiveness of the algorithm.