Single Image Super-resolution Reconstruction Based On Image Self-similarity

Image super-resolution is an important part of image processing and computer vision. In real life, it is widely used in many applications such as high-definition video, image compression and image restoration. The target of image super-resolution is to restore high-resolution images from a single(or multiple) low-resolution image(s). This paper focuses on restoring a high-resolution image from a single image. In comparison to the case of using multiple images, this problem is more uncertain. Hence, a strong a priori information is necessary to make it feasible. Image edges and multi-scale image similarity prove to be two types of promising a priori information. In recent years, many scholars have conducted in-depth research in this aspect, and many advanced super-resolution reconstruction methods are proposed. For this reason, we do the following two studies in this thesis.There are mainly two drawbacks lying in the existing super-resolution reconstruction methods based on the edge sharpening prior. First, these methods can achieve satisfactory performance only if highly accurate edge localization is guaranteed. Second, these methods usually ignore the multi-scale image similarity. To solve this problem, we propose a single image super-resolution learning method based on cross-scale(a special case of multi-scale) to learn the edge guidance. The core idea of this method is to provide a cross-scale learning framework of edge-directed sharpening function. First, the proposed method learns the sharpening function from a low and lower resolution image pair based on the similarities across different scales. Then, the sharpening function is applied to the high and low resolution image pair to estimate the sharpening edge prior of the high resolution image. Finally, we can restore the high resolution image by combining the edge-directed super-resolution framework and the estimated sharpening edges.For some existing learning based methods, they mainly learn the relationship between images from the existing training samples. If some type of images we want to reconstruct do not exist in the training samples, there is tend to be errors in the restoration results obtained by using these training samples. For example, the model trained from natural images is infeasible to be applied to other types of images, such as medical or remote sensing images. For this problem, we propose a super-resolution algorithm based on self-similarity regression. First, we propose an algorithm to generate multi-scale self-similar samples. Specifically, we perform continuous down sampling on a sample image to match it and its down sampled version as high and low resolution image pair. In this way, we obtain the high and low resolution image block pairs, which are used to train the regression model(a learning model). Meanwhile, inspired by the traditional linear regression, we propose a new bilinear regression model to learn the transformation matrix, which is used to transform the low resolution image blocks to high resolution ones.Based on the results of a large number of comparative experiments(including both visual and numerical), the feasibility of the two proposed super-resolution reconstruction methods is verified.