Image debanding using cross-scale invertible networks with banded deformable convolutions (Neural Networks 2025)

Description

A plug-and-play banded deformable convolutional layer significantly reduces the number of learnable parameters compared to standard deformable convolutional layers, providing a more efficient and stable alternative for training.

The receptive field at each position is modulated by a rotation scalar, a scaling factor, and an optional re-scale factor, resulting in a banded shape as follows:

For steady training, initialize the scaling factor at each postion as 1 and the rotation angle at each position as 0 degree.

The implementation is fully based on PyTorch, making it easy to integrate and use. Feel free to give it a try!

How to use

• The module supports convolutional operations by utilizing a non-square arrangement of sampling points, allowing the kernel_size to be specified as either an integer or a tuple.

• min_sscale/max_sscale: A pair of threshold for the scaling factor. If you would like to obtain a more flexible network, setting both of them as None is recommanded. Otherwise, pls treat them as hyper-parameters. Recommand to set min_sscale as 0.5 while setting max_sscale as 3.

• isRescale: (bool) default to True.

  • For non-local perception capacity: True is recommend. In theory, the magnification factor for the sampling interval can range from $[1, \infty]$.
  • For stable training: False is recommend.

• init_angle: default to 0, usually can be set to [0, 30, 45, 90]. Initialize the rotation angle with the given value.

Requirements

torch>=1.9.0
torchvision>=0.10.0

Citation

@article{quan2025image,
  title={Image debanding using cross-scale invertible networks with banded deformable convolutions},
  author={Quan, Yuhui and He, Xuyi and Xu, Ruotao and Xu, Yong and Ji, Hui},
  journal={Neural Networks},
  volume={187},
  pages={107270},
  year={2025},
  publisher={Elsevier}
}