Neighbor Embedding Based Image Super-resolution Reconstruction

In everyday life and actual production, images have become important information carrier of human perception environment. High quality images can more accurately reflect the content of the scene and have a lot to help the understanding of follow-up and decision-making. However, in the practical imaging process, due to the limitations of the degraded factors such as motion burring, fuzzy, down-sampling, and noising, it is difficult to obtain high resolution images or image sequences, which causes many difficulities for image processing, analysis, and understanding, leading to an obstacle in correctly understanding and perception of the laws of the objective world. Therefore, it is challenging to increase the spatial resolution of an image and to improve its quality. With image super-resolution technique, it is one of the effective methods to solve these problems. It can recover high resolution image for using signal processing technology and is a low cost of super-resolution technique. Accordingly, the SR technique has its wide application in many military, medical and civilian control, remote sensing imaging, computer vision, pattern recognition and so on.For the challenging problems of SR reconstruction, this thesis makes a deep research on example based SR methods, in which the philosophy of popular learning, neighbor embedding, low-rank matrix decomposition, joint learning. The main contributions are the following:1. An example-based SR method is proposed based low-rank neighbor embedding. To order to reduce the inconsistencies of the neighborhood relationship for the low resolution images and high resolution images. First of all, it improves the feature extraction method of the low resolution images; Then, it use the low rank matrix decomposition algorithm to enhance consistency mapping relations between the low resolution and high resolution image. Finally, make the choice of neighborhood and estimate the mapping relationship between the low resolution and high resolution image patches for their low rank component, and introduce a new method of the weight values, effectively improve the quality of image super-resolution reconstruction.2. An example-based SR method is proposed based Norm LV feature and sparse neighbor embedding. First, considering the traditional NE algorithm’s neighbor selection inaccurate for unreasonable features representation, this thesis constructs the new Norm LV enhanced feature to neighbor embedding. Then, to overcome the limitation of neighbor selection method and the reconstruction, an example-based SR method that is based on sparse neighbor embedding. Finally, the consistency prior and the global reconstruction constraint are applied to further enhance the quality of the initial SR estimate.3. To target the problem that the neighborhood relationship between LR image patches And the corresponding HR image patches in the traditional neighbor embedding methods cannot be perfectly preserved, an example-based SR method is presented by using joint learning sparse neighbor embedding. First, a joint learning is applied to train couple constraint and LR image and HR image projection matrices simultaneously to construct a unified feature subspace; then the low rank matrix decomposition algorithm is performed for getting the rank component. Finally, the sparse neighbor embedding are applied in the unified feature subspace.