Resources

• Implementations of SRDRM and SRDRM-GAN for underwater image super-resolution
• Simplified implementation of SRGAN, ESRGAN, EDSRGAN, ResNetSR, SRCNN, and DSRCNN
• Implementation: TensorFlow >= 1.11.0, Keras >= 2.2, and Python 2.7

Single Image Super-Resolution (SISR)	Color and sharpness

Pointers

• Paper: soon
• USR-248 dataset: http://irvlab.cs.umn.edu/resources/usr-248-dataset
• Bibliography entry for citation:

   article{islam2019fast,
       title={Underwater Image Super-Resolution using Deep Residual Multipliers},
       author={Islam, Md Jahidul and Enan, Sadman Sakib and Luo, Peigen and Sattar, Junaed},
       journal={arXiv preprint arXiv:x.y},
       year={2019}
   }
  

• Video demo: https://youtu.be/qOLZVgrxCwE

2x SISR performance	4x SISR performance

Usage

• Download the data, setup data-paths in the training scripts
• Use the individual scripts for training 2x, 4x, 8x SISR models
  • train-GAN-nx.py: SRDRM-GAN, SRGAN, ESRGAN, EDSRGAN
  • train-generative-models-nx.py: SRDRM, ResNetSR, SRCNN, DSRCNN
  • Checkpoints: checkpoints/dataset-name/model-name/
  • Samples: images/dataset-name/model-name/
• Use the test-scripts for evaluating different models
  • A few test images: data/test/ (ground-truth: high_res)
  • Output: data/output/
• A few saved models are provided in checkpoints/saved/
• Use the measure.py for quantitative analysis based on UIQM, SSIM, and PSNR

Constraints and Challenges

• Trade-offs between performance and running time. Requirements:
  • Running time >= 5 FPS on Jetson-TX2
  • Model size <= 12MB (no quantization)
• Challenges
  • Performance for 8x models
  • Inconsistent coloring, infrequent noisy patches

Acknowledgements

• https://github.com/wandb/superres
• https://github.com/david-gpu/srez
• https://github.com/Mulns/SuperSR
• https://github.com/tensorlayer/srgan
• https://github.com/icpm/super-resolution
• https://github.com/alexjc/neural-enhance
• https://github.com/jiny2001/dcscn-super-resolution
• https://github.com/titu1994/Image-Super-Resolution
• https://github.com/PacktPublishing/Advanced-Deep-Learning-with-Keras