Deep learning based super-resolution (SR) is one of the most actively studied areas of computer vision. However, many of these studies are conducted on a supervised manner, requiring a large amount of data. There are several problems with this. First, the ground truth and input pair of the dataset is made using only a specific procedure, usually the bicubic downsampling. As a result, supervised SR works well only for these images, introducing a second problem where the model can not work well for test images not found in the training distribution. Based on this, it is difficult to say that these methods indeed is super-resolution for raw images found in the real world. In this paper, we introduce a novel network named Zero-shot Super-Resolution with Generative Adversarial Network (ZSRGAN) for real world image SR, which needs only one test image and does not rely on any other external datasets. Unlike existing methods, we propose optimization on the perceptual aspect as well as reconstruction of pixel units using zero-shot SR method. Therefore, through our proposed model, a real world image can be super resolved with the best perceptual quality without any information from additional datasets.
First, the sample data(Degraded Set5) already are placed in <ZSRGAN_path>/datasets/MySet5
The results will save in <ZSRGAN_path>/experiments/
python train.py --name <save_result_path>
You can find additional dataset from Here provided by MZSR (CVPR 2020)
First, put your data files in <ZSRGAN_path>/datasets/
The results will save in <ZSRGAN_path>/experiments/
python train.py --name <save_result_path> --dataset <name_of_your_dataset> --GT_path <HR_folder_in_your_dataset> --LR_path <LR_folder_in_your_dataset>
Our project was based on ZSSR (CVPR 2018) and the data was taken from MZSR (CVPR 2020).