A PyTorch implementation of SRGAN based on the paper Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
- Anaconda
- PyTorch
conda install pytorch torchvision -c pytorch
- tqdm
pip install tqdm
- opencv
conda install -c conda-forge opencv
The train and val datasets are sampled from VOC2012.
Train dataset has 16700 images and Val dataset has 425 images.
Download the datasets from here, and then extract it into data directory.
The test image dataset are sampled from
| Set 5 | Bevilacqua et al. BMVC 2012
| Set 14 | Zeyde et al. LNCS 2010
| BSD 100 | Martin et al. ICCV 2001
| Sun-Hays 80 | Sun and Hays ICCP 2012
| Urban 100 | Huang et al. CVPR 2015.
Download the image dataset from here, and then extract it into data directory.
The test video dataset are three trailers. Download the video dataset from here.
python train.py
optional arguments:
--crop_size training images crop size [default value is 88]
--upscale_factor super resolution upscale factor [default value is 4](choices:[2, 4, 8])
--g_trigger_threshold generator update trigger threshold [default value is 0.2](choices:[0.1, 0.2, 0.3, 0.4, 0.5])
--g_update_number generator update number [default value is 2](choices:[1, 2, 3, 4, 5])
--num_epochs train epoch number [default value is 100]
The output val super resolution images are on training_results directory.
python test_benchmark.py
optional arguments:
--upscale_factor super resolution upscale factor [default value is 4]
--model_name generator model epoch name [default value is netG_epoch_4_100.pth]
The output super resolution images are on benchmark_results directory.
python test_image.py
optional arguments:
--upscale_factor super resolution upscale factor [default value is 4]
--image_name test low resolution image name
--model_name generator model epoch name [default value is netG_epoch_4_100.pth]
The output super resolution image are on the same directory.
python test_video.py
optional arguments:
--upscale_factor super resolution upscale factor [default value is 4]
--video_name test low resolution video name
--model_name generator model epoch name [default value is netG_epoch_4_100.pth]
The output super resolution video and compared video are on the same directory.
Upscale Factor = 2
Epochs with batch size of 64 takes ~2 minute 30 seconds on a NVIDIA GTX 1080Ti GPU.
Image Results
The left is bicubic interpolation image, the middle is high resolution image, and the right is super resolution image(output of the SRGAN).
- BSD100_070(PSNR:32.4517; SSIM:0.9191)
- Set14_005(PSNR:26.9171; SSIM:0.9119)
- Set14_013(PSNR:30.8040; SSIM:0.9651)
- Urban100_098(PSNR:24.3765; SSIM:0.7855)
Video Results
The left is bicubic interpolation video, the right is super resolution video(output of the SRGAN).
Upscale Factor = 4
Epochs with batch size of 64 takes ~4 minute 30 seconds on a NVIDIA GTX 1080Ti GPU.
Image Results
The left is bicubic interpolation image, the middle is high resolution image, and the right is super resolution image(output of the SRGAN).
- BSD100_035(PSNR:32.3980; SSIM:0.8512)
- Set14_011(PSNR:29.5944; SSIM:0.9044)
- Set14_014(PSNR:25.1299; SSIM:0.7406)
- Urban100_060(PSNR:20.7129; SSIM:0.5263)
Video Results
The left is bicubic interpolation video, the right is super resolution video(output of the SRGAN).
Upscale Factor = 8
Epochs with batch size of 64 takes ~3 minute 30 seconds on a NVIDIA GTX 1080Ti GPU.
Image Results
The left is bicubic interpolation image, the middle is high resolution image, and the right is super resolution image(output of the SRGAN).
- SunHays80_027(PSNR:29.4941; SSIM:0.8082)
- SunHays80_035(PSNR:32.1546; SSIM:0.8449)
- SunHays80_043(PSNR:30.9716; SSIM:0.8789)
- SunHays80_078(PSNR:31.9351; SSIM:0.8381)
Video Results
The left is bicubic interpolation video, the right is super resolution video(output of the SRGAN).
The complete test results could be downloaded from here.