The MWCNN is a network introduced by Pengju Liu et al. in "Multi-Level Wavelet-CNN for Image Restoration", CVPR 2018, and refined in "Multi-Level Wavelet Convolutional Neural Networks", IEEE Access June 2019. If you use this network, please cite their work appropriately.
The official implementation is available here in Matlab and here in Pytorch. A second version is available here in Pytorch.
The goal of this implementation in TensorFlow is to be easy to read and to adapt:
- all the code is in one file
- defaults are those from the paper
- there is no other imports than from TensorFlow
The only thing I am currently not happy about is the implementation of IWT.
Currently, it is very difficult to read because tensor slice assignment is
impossible as-is in TensorFlow.
See for example this SO question.
There are some differences in the 2 papers defining the MWCNN. We provide the 2 defaults and the ways to implement both slightly different architectures.
The Figures defining the implementations are 3. in each paper.
Some of these differences are acknowledged in the journal paper:
Compared to our previous work [24], we have made several improvements such as: (i) Instead of directly decomposing input images by DWT, we first use conv blocks to extract features from input, which is empirically shown to be beneficial for image restoration. (ii) In the 3rd hierarchical level, we use more feature maps to enhance feature representation.
You can implement the journal-style MWCNN simply with MWCNN().
To implement the conference-style MWCNN, you need to do the following:
from mwcnn import DEFAULT_N_FILTERS_PER_SCALE_CONF, DEFAULT_N_CONVS_PER_SCALE_CONF, MWCNN
model = Model(
n_filters_per_scale=DEFAULT_N_FILTERS_PER_SCALE_CONF,
n_convs_per_scale=DEFAULT_N_CONVS_PER_SCALE_CONF,
n_first_convs=0,
bn=True,
)The batch normalization is the only difference. It is present in the 2018 conference paper. It is absent in the 2019 journal paper.
In the 2018 conference paper:
Each layer of the CNN block is composed of convolution with 3 × 3 filters (Conv), batch normalization (BN), and rectified linear unit (ReLU) operations.
In the 2019 journal paper:
More concretely, each layer contains convolution with 3×3 filters (Conv), and rectified linear unit (ReLU) operations.
In the conference paper, there is a DWT operation as the first operation in the network. In the journal paper, there are first convolutions before the DWT.
In the conference paper, there are 4 convolution blocks per scale. In the journal paper, there are 3 convolution blocks per scale. The number of convolution blocks per scale are also different in the 2 papers.
I am welcoming any feedback under the form of GitHub issues or Pull requests.