Keras-WaveletTransform

This is an implementation of Wavelet Transform layer for denoising and classification from the paper "Multi-level Wavelet Convolutional Neural Networks" by Pengju Liu, Hongzhi Zhang, Wei Lian, and Wangmeng Zuo The paper can be found at : https://arxiv.org/abs/1907.03128

Denoising Application

Trained on DIV2K_train and tested on DIV2K_valid, CBSD68, Set12 and Urban100 Launch init.py with parameters

• -tr : Train dataset path (Train only)
• -t : Test dataset path
• -n : Noise level
• -s : Sliding window step (Test only)
• -lw : Weight path of the model to load (Test only)
• -a : Architecture of the model
• -m : Mode (Train or Test)

Example for training a model based on Unet with Noise level 15:

python3 init.py -tr DIV2K/DIV2K_train_HR/ -t DIV2K/DIV2K_valid_HR/ -n 15 -a unet -m train

Example for testing a pretrained model based on Wavelet with Noise level 50 and Sliding window step of 50p :

python3 init.py -t DIV2K/DIV2K_valid_HR/ -n 50 -a wavelet -m test -lw weights/DenoisingWavelet_50.h5 -s 50

Classification Application

Train and test a new model on Cifar100: Launch classificationCifar.py or classificationCifarWavelet.py

Benchmark

For Denoising Application only. Compares pretrained unet-based model to wavelet-based model following noise level (15, 25, 50). Computes mean SSIM and mean PSNR over a dataset. dataset_path and models in benchmark_list in the script can be changed for custom benchmark.