Non-negative Neighbor Embedding Based Single Image Super-resolution

Image is an important information media in modern life. No matter in everyday life or scientific research, high-resolution(HR) images are always desirable. The higher the resolution, the clearer edges and richer details images can provide. However, due to the physical conditions of the imaging system and adverse atmospheric condition, the practical imaging process is frequently prone to being degraded, resulting in blurred, down-sampled and noisy images, which are seldom satisfactory. Image super-resolution(SR) reconstruction is a software-based way to produce a high-resolution image from one or multiple low-resolution(LR) images with high feasibility. Therefore, research on image super-resolution algorithm has become an active topic in the field of image processing.The learning-based SR methods can recover the high-frequency component lost in the LR image by taking advantage of external training examples, thus achieving better results. Because of that, more and more scholars pay attention to this kind of method in recent years. The neighbor embedding based(NE-based) SR method is a representative solution. To obtain the HR image, the NE-based SR method embeds the neighborhood relationship in the LR space into the HR space directly. However, the NE-based SR method still deserves to be studied further in the following aspects: how to choose and process the external training examples, how to determine the number of nearest neighbors, how to speed the algorithm so as to meet the real-time requirement, how to suppress the artifacts caused by inappropriate choice of training examples. In view of the problems mentioned above, this thesis makes a deep analysis of NE-based SR method and achieves the following contributions:1. A single image SR method through non-negative neighbor embedding with pre-amplification is proposed. The NE-based SR method assumes that small LR and HR image patches can form manifolds with similar local geometry in the corresponding feature space. However, the assumption can’t hold in case of high magnification factor. What’s more, we can’t determine a generic number of nearest neighbors for different images. Therefore, we use pre-amplification technique to train the mapping relationship between high-resolution and middle-resolution images, leading to better neighborhood preservation of image patches. On the other hand, non-negative neighbor embedding method is proposed to solve the K selection problem.2. A non-negative neighbor embedding based single image SR method with non-local means regularization is proposed. The NE-based SR method is far too slow to meet actual demands. What’s worse, it is prone to producing artifacts if the training examples are improper. Therefore, we cluster the training image patches in the training phase and extract new feature for LR image patches so as to reduce the computation of finding nearest neighbors. On the other hand, we introduce a non-local means regularization term into the final reconstruction process by taking advantage of the self-similarity between natural image patches, which contributes to suppress artifacts.The two proposed methods can figure out the aforementioned problems effectively. Experimental results demonstrate that the proposed methods, which can achieve results with richer textures and sharper edges, are superior to the traditional methods in robustness and efficiency.