Feature Representation And Neighbor Embedding Based Image Super Resolution

Image super-resolution reconstruction(SR) is such a technology which can recovery high-resolution image from a single or several low-resolution counterparts, and it has developed to be one of research hotspots in computer vision field in recent years.Neighbor-embedding-based SR(NESR) is a popular group method of SR, and its core idea is to seek embedding manifold through a group of training samples, and realize the reconstruction of high resolution image patches with the assumption that small patches in the Low-Resolution(LR) and High-Resolution(HR) images form manifolds with similar local geometry in two distinct spaces(called manifold consistency assumption).However, the manifold consistency assumption is hard to satisfied for real-world images,and a great number of complex images are more likely to distribute on multiple manifolds structures, both of which will lead to inaccurate manifold embedding and have bad effect on the accuracy of reconstruction.Aiming at the above problems, this dissertation mainly researches on feature representation and neighborhood selection in neighbor-embedding-based SR methods,and the detail works are as follows:1. Maximal linear Patch based Neighbors Embedding SR. For the problems of feature selection, nonlinear manifold and neighborhood selection in NESR. Firstly, we extract the mid-frequency(MF) and high-frequency(HF) component as features and then Hierarchical Divisive Clustering algorithm will be conducted on the nonlinear manifold of training set for multiple maximal linear patch(MLP); Secondly, we select the most similar MLP for each test patch and considering the existing differences between image patches, we classify the test patch into edge or non-edge one, and then distinct selection method will be applied within its K nearest neighbors for final embedding neighbors.Finally, the HR reconstruction image can be achieved via linearly embedding those corresponding HF neighbors.2. Sparse Multi-manifolds Embedding for SR. For the problem of single-manifold assumption in NESR and fixed neighborhood size in NESR, the proposed method employs multi-manifold attribution assumption for image patches. Firstly we applyclustering method to construct different groups of training subsets. Then we match each LR input to its nearest training subset, and find appropriate number of neighbors from the same manifold by solving a constraint sparse optimization problem, which is more satisfy the reality.3. Single Image Super-resolution via Sparse Representative Features and Manifold Embedding. The proposed method has applied machine learning method for feature learning. A hierarchical support vector machine network is advanced to learn representative features for both training and test Low-Resolution(LR) image patches,and it is believed to break the limitation for ordinary feature extraction and obtain richer feature information as well. Then the learned features will be used in sparse manifold embedding.The above new methods have been simulated and though a series of experiments, the experimental results have verified the feasibility and effectiveness of this proposed method for natural images. Moreover, the proposed methods in this thesis can achieve competitive SR quality compared with other state-of-the-art baselines.