Dilation-based Blind-Spot Convolutional Neural Networks
TensorFlow 2.X implementation of the Blind-Spot Neural Network Architecture used in the paper Efficient Blind-Spot Neural Network Architecture for Image Denoising.
The principal building block is a new 2D convolutional layer with a blind-spot in the centre of its kernel. To preserve the blind-spot property throughout the entire network, blind-spot convolutions are progressively dilated as visualized in the graph below:
- tensorflow >= 2.0
- notebook, matplotlib (for visualisation of the receptive field)
If you find this work useful, please cite us:
@inproceedings{honzatko2020efficient,
title={Efficient Blind-Spot Neural Network Architecture for Image Denoising},
author={Honz{\'a}tko, David and Bigdeli, Siavash A and T{\"u}retken, Engin and Dunbar, L Andrea},
booktitle={2020 7th Swiss Conference on Data Science (SDS)},
pages={59--60},
year={2020},
organization={IEEE}
}