Block-matching based single model denoiser using Keras and TensorFlow

Meisam Rakhshanfar

This is an implementation of block-matching CNN based image denoiser BMCNN using Python 3, Keras, and TensorFlow. This work is similar to IRCNN. The differences are:

1. Prior to denoising, a block matching algorithm searches for similar blocks. For each 4x4 block, 4 similar blocks are found. Thus, an image with 5 channels (one original and 4 similar blocks) are fed into the network.

2. Possibility to use a lighter network. Since the block matching already does some of the processing we can use a simpler network with a fewer number of filters (fs = 24).

3. Unlike most of CNN based denoisers that for noise sigma specific weights are required, BMCNN uses single model (fixed weight) for all noise levels.

The repository includes:

• Source code of BMCNN.
• Block-Matching code
• Pre-trained weights BMCNN
• Jupyter notebooks to visualize the denoising results

Getting Started

• demo.ipynb Is the fastest way to start. It shows an example of using a model pre-trained for variation of noise levels. It includes code to run the denoiser on arbitrary images and different noise levels.

• model.py: This file contains the main BMCNN implementation.

• utils.py: This file contains some noise related functions.

Installation

1. Install dependencies

   pip3 install package [numpy, keras, skimage, ...]

2. Clone this repository

3. Run setup from the libs directory

   python3 setup.py install