Abstract
In this thesis, we propose a novel single-image super-resolution system that takes only one low-resolution input image under a learning-based image fusion framework. Although image super-resolution problem has been studied for decades, there is no such an approach that can work well for all different types of images under diverse blur levels, thus limiting most of the state-of-the-art approaches from practical usage. Therefore, we design a learning-based fusion system that integrates several representative image super-resolution approaches, including the interpolation-based, exemplar-based and reconstruction-based methods, for aggregating their advantages to obtain an adaptively fused image super-resolution result. The proposed approach is decomposed into two principal steps: initial high-resolution image estimation by different methods and adaptive image fusion with learning-based weighted integration. To start with, the initial high-resolution images are estimated by using three representative image scaling approaches; namely, the bicubic interpolation, reconstruction-based and exemplar-based algorithms. Subsequently, a learning-based approach is applied to build a weight table for the adaptive combination of the upscaled images in the patch based manner to obtain the optimal reconstruction. The weight table was learned from an external image dataset by using random projection trees for selecting a number of anchor points and linear polynomial fitting for each group associated with an anchor point. With the learned weight table, we can reconstruct a robust and superior high-resolution image by locally adaptive integration of the three initial upscaled images estimated by three distinct approaches. Our experiments demonstrate the high-quality image super-resolution results by using the proposed learning-based fusion algorithm. The proposed algorithm outperforms the competing single-image super-resolution algorithms through experimental comparisons on benchmarking images based on objective image metrics as well as subjective user study.