Enhancing low resolution images by applying deep network to produce high resolutions images.
- k3n64s1 this means kernel 3, channels 64 and strides 1.
- We downsample HR images to get LR images
- Generator is first trained for LR to SR generation, which has to be pixelwise close to HR (high PSNR)
- Discriminator is trained to distinguish SR & HR images, loss of which is backpropagated to train Generator further
- This game theoretic process will train generator to generate SR images that are indistinguishable from HR
- DIV2K and Flickr2K datasest (jointly DIV2F) for training
- Set5, Set14, BSDS100, Urban100, Manga109 for testing purposes
- main.py : Contains all running procedures
- myutils.py : Custom keras layers & function implementations
- models_collection.py : File containing all architecture definitions for fast protyping on changes
- Data reading & LR generation, with support for heterogenous & multiformat images
- Disk batching for control over memory usage by limited amount of data feeding
- Model as layer, dictionary based arguments & non-comparative model training
- GAN model with sub-models, and configuring layers for model API
- Model parser function for easier prototyping
- PixelShuffle & DeShuffle keras layers
- Learnable Weighted sum keras layer
- Plotting functions for Training progress
- Saving & Training resuming tricks for composite models
- MIX loss function with its Adaptive variant
- Attention Mechanism in GAN
- Data augmentation Class for Multi I/O keras models
- Multi-Scale Networks
- Network Interpolation
- Hybrid Padding for Testing on very large images
- Bicubic UpSampling layer
- Residual learning helps a lot in training of deep neural networks
- Perceptual loss can be useful for domains like Image Style Transfer and Photo Realistic Style Transfer also.
- Data collection is an import aspect, using images with high quality data, quality above quantity, as we have augmentation.
| DataSet Name | Minimum | Maximum | Mean | Median | Standard Deviation |
|---|---|---|---|---|---|
| DIV2K | 0.1340 | 3.6828 | 1.3501 | 1.1362 | 0.7683 |
| Set5 | 0.6227 | 4.0698 | 2.1655 | 1.9665 | 1.3567 |
| Set14 | 0.3126 | 3.3328 | 1.4685 | 1.2384 | 0.9801 |
| BSDS100 | 0.0241 | 4.5538 | 1.2028 | 1.0678 | 0.8341 |
| Urban100 | 0.3071 | 6.3944 | 1.6101 | 1.2864 | 1.0600 |
| Manga109 | 0.4461 | 6.0257 | 3.0673 | 3.0346 | 0.9606 |
| DataSet Name | Minimum | Maximum | Mean | Median | Standard Deviation |
|---|---|---|---|---|---|
| DIV2K | 0.4961 | 5.2995 | 1.6905 | 1.4589 | 1.0193 |
| Set5 | 0.8236 | 4.4487 | 2.6640 | 3.0197 | 1.4554 |
| Set14 | 0.4086 | 3.7932 | 1.7533 | 1.5810 | 1.0708 |
| BSDS100 | 0.1496 | 5.6223 | 1.4077 | 1.2298 | 0.9561 |
| Urban100 | 0.3184 | 6.7475 | 1.8952 | 1.5763 | 1.1998 |
| Manga109 | 1.2940 | 6.6388 | 3.8610 | 3.7790 | 1.1522 |
- Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
- ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks
- Perceptual Losses for Real-Time Style Transfer and Super-Resolution
- Efficient Super Resolution For Large-Scale Images Using Attentional GAN
- Getting started with the Keras functional API
- VGG usage in keras
- Various GANS implemented in Keras including SRGAN
- VGG loss help
- Improved Techniques for Training GANs
- Data Augmentation using Multiple Inputs
- Data Augmentation using Multiple Outputs
- https://github.com/titu1994/Super-Resolution-using-Generative-Adversarial-Networks
- https://github.com/brade31919/SRGAN-tensorflow
- https://github.com/MathiasGruber/SRGAN-Keras
- https://github.com/tensorlayer/srgan
You will need the following to run the above:
- Version of numpy 1.16.4
- Version of matplotlib 3.0.3
- Version of tensorflow 1.14.0
- Version of h5py 2.8.0
- Version of keras 2.2.5
- Version of PIL 4.3.0
Essential Hardware Requirement: A decent NVIDIA GPU 😜