Generalization ability

[daifeng2016]

Thank you for sharing the exciting work. I have three questions:

1. In traditional blind-SR setting, only LR images are provided such as kernelGAN/ZSSR, while in your work HR images are used, and the network are trained using synthesized LR images in a supervised setting. I am wondering if it is blindSR actually.
2. Although multiple Gaussian blur/noise are added when generating LR images, it is still far from real-world degradation process. I am wondering if the model generalizes well to the real-world scenarios, where the degradation might not be Gaussian blur or noise.
3. Can your model work in real world unpaired LR/HR setting,where no systhesized images are used but unpaired LR and HR images are provided?

[LongguangWang]

Hi @daifeng2016, thanks for your interest in our work.

1. First, blind SR is defined as super-resolving an LR image without knowing its degradation model. During test phase, our DASR takes only an LR image as input to produce an HR image as ZSSR does. Thus, we believe our network can be considered as a blind SR method. Second, it should be noted that the learning of degradation representation in our DASR is unsupervised since the groundtruth degradation is not used as supervision.
2. It is true that Gaussian blur and noises cannot cover all possible real-world degradations. In our experiments, our DASR generalizes well on several real-world images (Fig. 8 in the main text and FIg. VII in the supp). However, when the real degradation is far from the ones used during training, a notable performance drop can also be observed (Fig. III in the supp).
3. In this work, we focus on blind SR using paired images and our DASR cannot be directly trained on unpaired data. We will investigate unpaired blind SR in our furture works.