joshbaradia/esrgan_240000

We propose an attention-guided UNet leveraging spatial/channel attention and feature pyramid integration to correct color and restore details in underwater and also Enhansing the resolution of any Underwater Image using GANs.

Code will be updated once the paper is published :)

✨UFPN

🐳UFPN Training

Will be updated once the paper gets published. Our paper will be published in IEEE Xplore Library

✨ESRGAN

Enhancing images upto 4x times. Using UFPN output images as input to our ESRGAN model.

🐳ESRGAN Training

Make sure that all the dependencies are satisfield use requirements.txt file to load the required dependencies.

pip install -r requirements.txt

• Use the file /options/train_test_dataset_4x.yml to Fine-Tune the ESRGAN model which is RRDB_ESRGAN_x4.pth model.
• The new fine-tuned model gets saved.
• You can find the model in experiments/pretrained_models/ folder.

🐳Testing UFPN+ESRGAN

Go to https://github.com/JoshBaradia/ML-PIE

Using our best UFPN model checkpoint_trial_37.pth and net_g_240000.pth

• Run the test_test.py python script. The UFPN model takes input, which are the Low Resolution-Degraded images from /testrun_enet folder and the script saves it in /inputs folder.
• Then our testing script inference_realesrgan.py uses our ESRGAN model, takes input images from the /inputs folder which now has enhanced images and saves it in /results folder.
• Thus the /results folder contains enhanced-super-resolution images.

python test_test.py;  python inference_realesrgan.py --model_path experiments/pretrained_models/net_g_240000.pth --input inputs

The "experiments” folder and "checkpoint_trial_37” has been uploaded on huggingface.co as the files were larger. The model can be downloaded via the following link, access for the model will be automatically granted once requested! https://huggingface.co/joshbaradia/esrgan_240000