Single Image Super-Resolution via Deep Dense Network

Publication Type:
Issue Date:
Full metadata record
Image Super-Resolution (SR) is a research field of computer vision, which enhances the resolution of an imaging system. The need for high resolution is common in computer vision applications for better performance in pattern recognition and analysis of images. However, recovering of the HR image from LR image is a highly ill-posed problem. In this thesis, the image SR problem is solved from three aspects with deep dense network models, including improving reconstruction accuracy, optimizing model training-time memory consumption, and extending effective SR scale ranges. Chapter 1 introduces the importance of image SR reconstruction and summarizes the challenges of image SR problem. Chapter 2 reviews the existing image SR methods, analyses their limitations and explains some related fundamental theories. Chapter 3 proposes a bi-dense model to improve image SR performance based on the dense connections for feature reuse. The bi-dense network does not only reuse local feature layers in the dense block, but also reuses the block information in the network to archive excellent performance with a moderate number of parameters. Chapter 4 evaluates the memory consumption of the vanilla dense model for image SR. For solving this problem, we introduce shared memory strategy into image SR by proposing a memory-optimized deep dense network. Chapter 5 discovers most of the deep SR methods are inefficient or impractical for generating SR of any scale factor, and proposes a novel Any-Scale Deep Network (ASDN), which requires few training scales to achieve one unified network for any-scale SR. In order to design such a powerful network architecture, we propose Laplacian Frequency Representation to predict SR results of the small ratio range and Recursive Deployment for SR of any larger scale. In this way, the required training data and update periods are substantially decreased to optimize the any-scale SR network. All these algorithms are aimed to solve the single image SR problem. These algorithms are tested on many public datasets and the results on those datasets demonstrate superior performance of our approach over the state-of-the-art methods and validate the effectiveness and correctness of our methods.
Please use this identifier to cite or link to this item: