Chapter 1 - Video enhancement and super-resolution

Abstract

This study expects to enhance visual effects or visual features by exploring a novel Image Enhancement (IE) mechanism for low-resolution and low-quality video images. Consequently, an Attention Mechanism (AM)-improved General Adversarial Network (GAN) is proposed based on Deep Learning (DL), namely, the Least Squares GAN (LSGAN) model, which can deal with image quality degradation and Motion Blur (MB) problems. Specifically, the high-order gated AM is adopted to automatically learn different attention levels to realize the super-resolution and motion deblurring of the synthetic image. The results prove that the proposed model can better retain low-frequency content and deal with large-scale motion based on the fusion of nonlocal features and the time-domain information. The test of MB data sets shows that the proposed model can restore multidegradation in most cases; meanwhile, the proposed image quality restoration method is better on the synthetic dataset DIVerse 2K (div2k) and the real dataset GoPro (PSNR = 27.8 db and SSIM = 0.85), and there is no significant difference. Therefore this study uses multiframe fusion and nonlocal modules to process complex motion images, and has gained a good video IE effect.