Efficient Scale-Recurrent Network Using Generative Adversarial Network for Image Deblurring
Wei-Hsiang Hsu, Chih‐Wei Tang · 2022
Ubiquitous blurry images degrade viewing experiences and performance of video analysis. To be applicable to consumer electronics, the large amount of parameters and high computational load of deblurring network have to be avoided. The existing SRN+ has a small number of parameters (3.9M) and comparable performance. To improve the quality of outputs of a deblurring network (e.g., SRN+) without changing its architecture, this paper proposes the pseudo label based order task for training the discriminator. The proposed funnel soft label further reduces the problem of vanishing gradient during training SRN+ (generator), and the adversarial loss combined with weighted scale-level losses improves quality of deblurring. For GoPro dataset, the proposed scheme outperforms the light version of the state-of-the-art MPRNet in PSNR (+1dB) and number of parameters (70%).