Multi-Scale Laplacian Super Resolution Network
This is an Keras implementation based on the following research with a few tweaks,
Fast and Accurate Image Super-Resolution with Deep Laplacian Pyramid Networks.
The model has multi-scale training, shared-source skip connection and shared parametes for each upscale stage. With these properties, it achieves a depth of ~40 layers with just ~120,000 trainable parameters (32 filters for each 3x3 convolutional layer). Because of this, it is suitable for small datasets.
The model is defined in msLapSRN_model.py. Current implementation increases the resolution by 16 times (4x for width and height).
The data is read in with a Sequence class for training with the fit_generator function. The template is in dataset.py. The data should be stored in the HDF5 format with the following structure
- train/data --- low resolution images
- train/label_x2 --- 2x images
- train/label_x4 --- 4x images
Finally, the L1 Charbonnier loss and the training pipeline is defined in train.py. Once the training data is structured as described above, the model can be trained by just
python train.py