GAN-Image-Super-Resolution

End-to-End Keras implementation on top of Tensorflow 2.3

• Part of a Biometric-Backdoors research project.
• Image Super-Resolution Using a Generative Adversarial Network (SR-GAN)

SR-GAN Architecture

Train on your own Data

Fork the repository.

Dump the HR(High-Resolution) images under Data/HR/ and LR(Low-Resolution) images under Data/LR/.

Make sure about

    HR Images (Totall sample, 96*4, 96*4, 3)
    LR Images (Totall sample, 96, 96, 3)

Change tot_sample=Totall sample in traning data and Run the following code in current directory for TRANING.

    python train.py

Model will get saved in checkpoint folder in running EPOCHS.

Results

• Note:
  1. LRI- Low Resolution Input (96x96x3)
  2. HRP- High Resolution Prediction (384x384x3)
  3. RHR- Reference High Resolution Image(384x384x3)
• Results obtained with batch_size=5, Traning sample= 100 and Epoch=200

------------- LRI--------------------------- HRP ------------------------ RHR --------------

Data

• Download data from DIV2K - bicubic downscaling x4 competition dataset.
• Other direct links: test_LR_bicubic_X4, train_HR, train_LR_bicubic_X4, valid_HR, valid_LR_bicubic_X4.

References

1. SR-GAN- https://github.com/tensorlayer/srgan
2. Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network- https://arxiv.org/pdf/1609.04802.pdf
3. SUPER-RESOLUTION WITH DEEP CONVOLUTIONAL SUFFICIENT STATISTICS- https://arxiv.org/pdf/1511.05666.pdf