Date of Graduation
Statler College of Engineering and Mineral Resources
Lane Department of Computer Science and Electrical Engineering
Image demosaicing, image super-resolution and video super-resolution are three important tasks in color imaging pipeline. Demosaicing deals with the recovery of missing color information and generation of full-resolution color images from so-called Color filter Array (CFA) such as Bayer pattern. Image super-resolution aims at increasing the spatial resolution and enhance important structures (e.g., edges and textures) in super-resolved images. Both spatial and temporal dependency are important to the task of video super-resolution, which has received increasingly more attention in recent years. Traditional solutions to these three low-level vision tasks lack generalization capability especially for real-world data. Recently, deep learning methods have achieved great success in vision problems including image demosaicing and image/video super-resolution. Conceptually similar to adaptation in model-based approaches, attention has received increasing more usage in deep learning recently. As a tool to reallocate limited computational resources based on the importance of informative components, attention mechanism which includes channel attention, spatial attention, non-local attention, etc. has found successful applications in both highlevel and low-level vision tasks. However, to the best of our knowledge, 1) most approaches independently studied super-resolution and demosaicing; little is known about the potential benefit of formulating a joint demosaicing and super-resolution (JDSR) problem; 2) attention mechanism has not been studied for spectral channels of color images in the open literature; 3) current approaches for video super-resolution implement deformable convolution based frame alignment methods and naive spatial attention mechanism. How to exploit attention mechanism in spectral and temporal domains sets up the stage for the research in this dissertation. In this dissertation, we conduct a systematic study about those two issues and make the following contributions: 1) we propose a spatial color attention network (SCAN) designed to jointly exploit the spatial and spectral dependency within color images for single image super-resolution (SISR) problem. We present a spatial color attention module that calibrates important color information for individual color components from output feature maps of residual groups. Experimental results have shown that SCAN has achieved superior performance in terms of both subjective and objective qualities on the NTIRE2019 dataset; 2) we propose two competing end-to-end joint optimization solutions to the JDSR problem: Densely-Connected Squeeze-and-Excitation Residual Network (DSERN) vs. Residual-Dense Squeeze-and-Excitation Network (RDSEN). Experimental results have shown that an enhanced design RDSEN can significantly improve both subjective and objective performance over DSERN; 3) we propose a novel deep learning based framework, Deformable Kernel Spatial Attention Network (DKSAN) to super-resolve videos with a scale factor as large as 16 (the extreme SR situation). Thanks to newly designed Deformable Kernel Convolution Alignment (DKC Align) and Deformable Kernel Spatial Attention (DKSA) modules, DKSAN can get both better subjective and objective results when compared with the existing state-of-the-art approach enhanced deformable convolutional network (EDVR).
Xu, Xuan, "Video and Image Super-Resolution via Deep Learning with Attention Mechanism" (2020). Graduate Theses, Dissertations, and Problem Reports. 7678.