Similar Structure Is Based On Multi-scale Study Of A Single Image Super-resolution Reconstruction

In many digital image applications, image with high spatial resolution contains much information and is easy to be observed and identified. Image super resolution reconstruction is to recover a high resolution image from one image or multiple low resolution images with the same scene. The technique has been widely used in many applications including monitoring system, military reconnaissance, medical imaging and so on. Focusing on single image super resolution, this thesis systematically studied the methods of super resolution based on the iterative back projection approach and on the learned structure similarity. The description of our work is as follows:1) presented a modified iterative back projection framework. The algorithm improved the quality of initial amplified image, enhanced the contrast of the details for the intermediate iteration result. Experiments results proved that the modified approach could generate sharper edges and had a powerful denoising ability.2) proposed a reconstruction method by searching similar patches in multiple scales and weighing them adaptively. The reconstruction result based on learning is not natural, and the speed is slow. In this thesis, the learned high resolution blocks are modified and adaptively weighted according to the similarity of overlap regions inside adjacent blocks. Furthermore, this thesis introduced randomized patch-match technique for searching similar patches, which speed up the searching procedure. In addition, by decreasing the upsample factor of the adjacent two scales, the proposed algorithm reduced the number of similar patches and the searching space. Experimental results demonstrated the accuracy and efficiency of our proposed method.3) presented an adaptive segmentation algorithm to partition complex and disorderly texture area in images. On one hand, large upsample factor is limited using iterative back projection algorithm. On the other hand, super-resolution based on learning performed poorly in the complex and disorderly texture area, such as animal hair and lush foliage. Many false edges would be generated in the area. This thesis used IBP method to process the complex and disorderly texture area, and applied the improved multi-scale similar structure learning algorithm for the left regions. The experimental results validated our method.