Image-Super-Resolution-Using-GAN-SRGAN

This repo contains the project work carried out for the course Deep Learning in my B.Tech Final Year DA-IICT. It is the replication of the code in simpler terms available on GitHub.

1. Reference Paper Link: Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
2. Reference Code Link: Github

Dataset

You need to have the high resolution images for training. In this project, we have used images from DIV2K - bicubic downscaling x4 competition, so the hyper-paremeters are (like number of epochs) are seleted basic on that dataset, if you change a larger dataset you can reduce the number of epochs. Following are some links to download training andd test dataset:

1. Valid HR
2. Train HR
3. Valid LR
4. Train LR

Model/ Architecture

Pre-Trained Model

We have trained our model upto 100 epochs. Weights can be found from the Pre Trained Weights/ folder. You can fine tune the parameter and resume the training of the model. To load the model weights, follow the step shown below.

• Provide the path of your weights in the first cell variable G_weights_load and G_weights_load.
• Uncomment 2 lines i.e line no 16 and 17 in the training cell.
  • netG.load_state_dict(torch.load(G_weights_load))
  • netD.load_state_dict(torch.load(D_weights_load))

Run

• Open ipynb file in either google colab.
• Put the notebook on GPU mode.
• Change the path depending up on your file structure in the first cell.
• Run all cells.

Results

• Sample results are available in the Results/sr/ folder.

Other Collaborators

1. Ruchit Vithani
2. Bhargey Mehta
3. Kushal Shah