Research On Image Super-resolution Reconstruction Algorithm Based On Neighbor Embedding

Image super-resolution (SR) reconstruction is the process of generating a high-resolution image by using one or multiple low-resolution ones. As a post processing technique, it improves image quality through the way of restoring missing details, which are lost in the acquisition process. And compared to the way of improving image quality with better imaging device hardware, image super resolution technique has a wide range of applications.Currently, existing super resolution methods can be broadly categorized into three categories:interpolated-based, reconstruction-based and learning-based methods. Compared with the former two methods, learning-based methods have better performance in image reconstruction effect and can effectively solve the problem of a decline in reconstruction image quality when sampling factors are large. Therefore, learning-based methods have become a hotspot in the field of image super resolution. This thesis mainly takes learning based methods to begin our study, which provides detailed analyses on neighbor embedding image reconstruction methods and the application of image multi-scale self-similarity in super-resolution reconstruction. Also some optimizations and extensions are made according to the existing problems in existing super-resolution image reconstruction algorithms to achieve an enhancement in image reconstruction effect.The main research contents of this article are as follows:At first, taking image super-resolution reconstruction algorithms with neighbor embedding scheme into consideration, we make a detailed analysis on feature representation for different image patches and weight compute for neighbors patches. Then based on the study in image multi-scale self-similarities, we propose a algorithm for single image reconstruction by processing low-resolution images to provide image pairs for training sets. And to improve the efficiency, we narrow the search spaces with in-place matching to search for local image patches. Experimental results show that the method is effective and reasonable.