This a slightly modified version of SRGAN model as it does not have the BatchNormalization layer as mentioned in the original paper. The training happens in two stages. The data preprocessing and metrics have been modified to support satellite reflectance data values (0-1) instead of traditional 0-255 range.
The model supports only 2x, 3x and 4x upscaling of TIF Images however it uses feature maps of VGG network in the final stage
AGRI
URBAN
- PSNR Training
First very deep ResNet architecture using the concept of GANs to form a perceptual loss function as measured by PSNR and structural similarity (SSIM) with our 16 blocks deep ResNet (SRResNet) optimized for MSE.
- GAN Training
Traditional SRGAN model is used with L1 Loss for training. The backend used is VGG-19
Sample data is available here. Standard TIF imagery (reflectance) is used for model building. The zip file should be extracted into data/ directory
Separate conda environment with Python >= 3.7 and tensorflow-gpu is recommended.
python setup.py install
- Configure File:
options\train\SRResNet_SRGAN\train_MSRResNet_x4.yml - Configure File:
options\train\SRResNet_SRGAN\train_MSRGAN_x4.yml
- Configure File:
options\test\SRResNet_SRGAN\test_MSRGAN_x4.yml
- Running the SRRSNET Model
python basicsr/train.py --options options\train\SRResNet_SRGAN\train_MSRResNet_x4.yml - Running the SRGAN Model
python basicsr/train.py --options options\train\SRResNet_SRGAN\train_MSRGAN_x4.yml - Inference from SRGAN Model
python basicsr/test.py --options options\test\SRResNet_SRGAN\test_MSRGAN_x4.yml
Inspired from https://github.com/xinntao/BasicSR