This repository contains the source code, data, and documentation for Onet, a novel O-shaped Twin U-Net architecture designed for unsupervised binary semantic segmentation in radar and remote sensing images. Source paper [onet2025] (preprint version) and the supplementary materials can be downloaded from this repository.

In marine radar domain, there are very few labeled datasets for training the deep neural networks. In this work, we propose an unsupervised model for binary semantic segmentation in marine radar images. Additionally, we demonstrate its efficiency for cloud detection in remote sensing images.

The basic idea behind this model is that, in marine radar images, the main distinction between the foreground and background is their intensity. A normalized image $X$ highlights the foreground, while its reverse, $1-X$, emphasizes the background. By feeding the network with the complementary input pair $(X, 1-X)$, the features of the foreground and background can be contrastively learned within the same feature space.

Considering that shallow convolutional networks capture local intensity variations, while deeper layers extract global features, it is beneficial to connect the shallow and deep layers directly to emphasize the influence of intensity on classification. This idea aligns well with the U-Net architecture, which uses skip connections to integrate features between the encoding and decoding layers.

Motivated by these insights, we utilize the initial layer for local feature expression and the final layer for global feature extraction in the U-Net network. The class probability maps are then generated by projecting the global features onto the local features at the corresponding locations. The segmentation probability is expected to heavily depend on the numerical distributions of the local features.

To strengthen this dependence, we propose to maximize the mutual information between local features and segmentation scores by raising its lower bound, the Jensen-Shannon Divergence (JSD) [mine2019], [infoseg2021]. To effectively handle the complementary input pairs, we propose a dual U-Net architecture which is called Onet in this paper due to its O-shaped structure.

_{Fig. 1 Architecture of Onet. Complementary input pairs, a symmetric structure and JSD loss encourage the network to learn distinctive features in binary semantic segmentation.}

Here are some reulsts demonstrating Onet's performance in taret segmenation in marine radar images and cloud detection in remote sensing images.

_{Fig. 2 Target segmentating comparsion in the rain clutter of the marine radar images.}

_{Fig. 3 Cloud detection in the remote sensing images.}

[mine2019] D. Hjelm, A. Fedorov, S. Lavoie-Marchildon, K. Grewal, P. Bachman, A. Trischler, and Y. Bengio, “Learning deep representations by mutual information estimation and maximization,” in International Conference on Learning Representations, 2019, pp. 9153–9176.

[infoseg2021] R. Harb and P. Kno ̈belreiter, “Infoseg: Unsupervised semantic image segmentation with mutual information maximization,” in German Conference on Pattern Recognition 2021.

[onet2025] Y. Zhou, H. Su, T. Wang and Q. Hu, "Onet: Twin U-Net Architecture for Unsupervised Binary Semantic Segmentation in Radar and Remote Sensing Images," in IEEE Transactions on Image Processing, vol. 34, pp. 2161-2172, 2025

@ARTICLE{ZhouTIP2025,
  author={Zhou, Yi and Su, Hang and Wang, Tian and Hu, Qing},
  journal={IEEE Transactions on Image Processing}, 
  title={Onet: Twin U-Net Architecture for Unsupervised Binary Semantic Segmentation in Radar and Remote Sensing Images}, 
  year={2025},
  volume={34},
  number={},
  pages={2161-2172},
}