Research On Self-Similarity Based Algorithm For Single Image Super-Resolution Reconstruction

Due to limitations of factors of hardware and environment, in many cases, the resolution of the image we collected is not ideal. If we improve the resolution of the image by improving a hardware device or to improve image acquisition environment, the cost may be high even impractical. So by using software methods to improve the image resolution becomes a practical approach. Super-resolution reconstruction technique is to use the digital image processing technology by using one single low-resolution image or multiple low-resolution images from the same scene to reconstruct the high frequency details of an image. The technology has broad application prospects in terms of video surveillance, medical imaging, and satellite remote sensing, and so on.This thesis focusing on the technology of learning-based single image super-resolution reconstruction, makes a systematic and deep study on the sparse-representation-based and dictionary learning algorithm, and proposes an algorithm using self-similarity and dictionary neighborhood for image reconstruction. The main contents of this thesis are:First, we make an overview on the popular super-resolution reconstruction algorithm. Since the traditional method based on interpolation and reconstruction have their own limitations, it is difficult to have a new breakthrough. So we focus on the study of learning-based super-resolution technology, and compare the various algorithms to help select the subsequent algorithm.Secondly, we make a deep discussion on the current popular sparse representation theory, and the discussion includes the selection of sample, training dictionary, and adjust parameters. Wherein, we make a key study on the selection of sparse parameters and establishing of dictionaries atoms.Finally, based on the previous research, we propose a super-resolution reconstruction algorithm based on the self-similarity of an image and dictionary neighborhood. The algorithm does not require external images as the training set, but to use the self-similarity of the in-scale or cross-scale of the input image itself, and to build a pyramid image, thereby obtaining a set of image pairs of low-resolution and its corresponding high-resolution. Then we train this set to get a dictionary and construct a dictionary neighborhood for every atom of the dictionary. We select the nearest dictionary neighborhood for every path of the input to reconstruct a super-resolution image. This algorithm can solve the problem of adaptive image well. Experimental results show that compared with the conventional method, the algorithm can obtain better reconstruction effect both in object data PSNR and personal eye visual effect, while the space requirements are also reduced.