Unofficial Tensorflow implementation of DBPN CVPR 2018.
Paper presentation is here.
- One of the Deep Super-resolution networks that learn representations of low-res inputs, and the non linear mapping to high-res output.
- This network exploits iterative up and down convolution layers thereby providing a negative feedback mechanism for projection errors at each stage.
- Stacking of each up and down sampling stage in this fasion is DBPN.
- Concatenation of all other previous feature layers with the current layer lead to the Dense version of DBPN. This is proven to have performance enhancements.
- We have added Perceptual Loss which is generally used for Realtime style transfer and super resolution.
- This loss emphasizes on optimization of high level features, learnt in pre trained networks (VGG16 used here, support can be extended), to be reconstructed by our network.
- We have used only Feature Reconstruction Loss of perceptual loss for training purposes. Style Transfer Loss is implemented though.
- Total Loss function
f: Non Linear Mapping (DBPN)
g_j: Pre-trained Network layer j
| Code | Utilities | Run Command |
|---|---|---|
| dbpn_lite.py | Implementation of DBPN with no dense connections. It has implementations of various scales 2x, 4x, 8x. It contains utility functions and model definitions. | |
| main.py | Main training program | python main.py --no_epoch 10000 --batch_size 8 --train_dir <path to train dataset> |
Training dataset for DBPN is only a folder full of HR images. LR down conversion is taken care by the input pipeline.
[Trained for limited number of iterations. Further improvements wrt to training can be done]
| Input LR | Predicted HR | True HR |
|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent