Related Experiment Video
Updated: Jul 14, 2026
10:25
Deep Learning-Based Segmentation of Cryo-Electron Tomograms
Published on: November 11, 2022
Onet: Twin U-Net Architecture for Unsupervised Binary Semantic Segmentation in Radar and Remote Sensing Images
Summary
This study introduces Onet, a novel deep learning model for unsupervised binary semantic segmentation. Onet effectively segments objects in cluttered images, improving marine radar target detection and remote sensing image analysis without needing labeled data.
Area of Science:
- Computer Vision
- Deep Learning
- Image Segmentation
Background:
- Object segmentation in single-channel images is challenging due to limited information and diverse targets.
- Existing methods struggle with cluttered backgrounds in marine radar, medical, and remote sensing imagery.
Purpose of the Study:
- To propose Onet, an O-shaped assembly of twin U-Nets, for unsupervised binary semantic segmentation.
- To achieve high-performance segmentation without requiring annotated labels.
Main Methods:
- Onet utilizes an intensity-complementary image pair for training.
- It maximizes Jensen-Shannon divergence (JSD) between localized features and class probability maps.
- The model leverages U-Net symmetry to enhance feature dependencies.
Main Results:
- Onet demonstrates superior performance in segmenting marine radar targets and remote sensing clouds.
- The model significantly improves the signal-to-noise ratio (SNR) of marine radar targets.
- Achieved state-of-the-art results in unsupervised segmentation tasks.
Conclusions:
- Onet offers a powerful solution for unsupervised semantic segmentation in challenging single-channel images.
- The method enhances target visibility in noisy environments like marine radar.
- Onet's open-source availability facilitates further research and application.
