Optimization of generative adversarial network based image super-resolution by using image mask
Article
Figures
Metrics
Preview PDF
Reference
Related
Cited by
Materials
Abstract:
Inferring high resolution image from single low resolution (LR) input is ill-posed and deep learning helps to some extent. The latest algorithms take the advantage of the Generative Adversarial Network (GAN) and present photo-realistic results by learning low/high resolution mappings from super resolution datasets. However, training of GANs can be hard and traditional GAN-based architectures often exhibit noise and texture distortion in their super-resolution (SR) results. In this paper, a mask-aided adversarial training strategy for current GAN-based SR frameworks is proposed. During training, mask module helps the discriminator by introducing additional perceptual quality information with generator’s outputs and the ground truth images. In experiment, three current state-of-the-art GAN-based SR models are selected and the mask module is integrated into their adversarial training. The improved mask-aided models yield better results in both quantitative and qualitative benchmarks than the original ones. Mask module only modifies GAN framework and thus is suitable for many GAN-based solutions for further improving the SR perceptual quality.
Keywords:
Project Supported:
Supported by the National Key Research and Development Program (2018YFB2100500) and National Natural Science Foundation of China (61976138, 61977047).
