Research On Super-Resolution Method

Super-resolution is a hot issue in computer vision and image processing, and has the potential to be widely used in many applications. This thesis focuses on the super-resolution reconstruction for various digital media, including image super-resolution, video super-resolution and hyperspectral image super-resolution. The main contributions of this thesis are summarized as follows.1. Propose a robust and fast method for image super-resolution based on sparse rep-resentation. This method formulate the super-resolution problem as a compressive sensing system, in which an over-complete dictionary is used to sparsely represent a low resolution image and generate a high resolution image. Then, the training images are adaptively con-structed by selecting the most correlative images based on scale-invariant feature transform (SIFT). The proposed method simultaneously ensures the quality and running time of the super-resolution.2. Propose two video super-resolution methods. The first method is a feature-guided variational optical flow approach, which uses optical flow warping and inpainting based on fast marching method (FMM). This method overcomes the block artifacts and considering the case of small structures with large displacement. Based on an adaptive superpixel-guided auto-regressive model, the second method achieves more accurate and robust super-resolution reconstruction by simultaneously exploiting the spatiotemporal correlations.3. Propose a simple and fast hyperspectral image super-resolution method, which can be widely used in various spectral images. The proposed method simultaneously exploits spatial and spectral correlations by using an accompanied color image. The spatial corre-lation is modeled by an auto-regressive model incorporated with a non-local mean (NLM) kernel. The spectral correlation is exploited by an adaptive warping method.The performances of our proposed methods are intensively evaluated with both simu-lations and real datasets.