Single Image Super-Resolution Reconstruction Based On Compressive Sensing

Taking full advantage of the existing images, super-resolution reconstruction(SRR) breaks through the resolution limitation of image by algorithm. Although the multi-frame image SRR theory on the basis of supplementary information has been consummate, but can’t solve the single image SRR problem without complement. The emerging compressive sensing(CS) theory explains the signal in sparse domain,and can solve reconstruction problem from low-dimensional data to high-dimensional data. Recently, the more meaningful single image SRR combined with CS theory has become a research hotspot.Currently, the existing SRR algorithm based on CS mainly focus on the simulation degraded image known the priori information. It’s difficult for the reconstruction result to surpass the resolution of the original image. To achieve the substantial increase of the original image, the main research and conclusions in the thesis are as follows:(1) Based on the realization and comparison of classical SRR algorithms,this thesis has implemented the training of DCT, MOD, K-SVD and Global dictionary with researching the construction method of sparse dictionary. By utilizing the four dicitonaries to sparse coding, the reconstruction results verify the feasibility of single image SRR based on CS. The PSNR and SSIM values show that training dicitionary can express the image more sparsely, and achieve the better reconstruction quality.(2) By training the coupled dicitionary of each component in YCb Cr and RGB color mode respectively, this thesis has designed a single color image SRR method based on CS. Compared with the SRR result only using Y component, the PSNR of overall reconstruction in YCb Cr has a slight increase, and more edge details can be rebulided in RGB color mode, but the reconstruction speed needs improving. In addition, natural image and remote sensing image are respectively selected for training coupled dictionary. This thesis has employed the above dictionaries to achieve single remote sensing image SRR. The edge extraction and PSNR results confirm the necessity of adaptive coupled dictionary.(3) Combined with appreciation and iteration,the thesis proposed an improved framework based on CS to realize the SRR of non-degenerate single color image and remote sensing image. The framework has been applied in single and multiple SRR in this thesis, and taken the edge extraction results, image information entropy and average gradient as the evaluation indexes to assess the framework. The result confirms the validity of the proposed framework.