Related Experiment Video
Updated: Aug 10, 2025
05:58
Using Rapid Serial Visual Presentation to Measure Set-Specific Capture, a Consequence of Distraction While Multitasking
Published on: August 29, 2018
8.9K
SAR Target Recognition with Limited Training Samples in Open Set Conditions
Xiangyu Zhou1, Yifan Zhang1, Di Liu1
1School of Software, Northwestern Polytechnical University, Xi'an 710129, China.
Sensors (Basel, Switzerland)
|February 11, 2023
Summary
This study introduces a novel open set recognition (OSR) method for synthetic aperture radar (SAR) target recognition, effectively handling unseen categories with limited training data. The approach achieves high accuracy for known targets and notable performance for unknown targets, offering valuable interpretability.
Area of Science:
- Remote Sensing
- Computer Vision
- Artificial Intelligence
Background:
- Collecting diverse training samples for all synthetic aperture radar (SAR) target categories is challenging.
- Open set recognition (OSR) is crucial for real-world scenarios where unseen categories may appear.
- Existing OSR methods for optical images or those requiring extensive data are often unsuitable for SAR due to data scarcity and unique characteristics.
Purpose of the Study:
- To propose a task-oriented OSR method specifically designed for SAR target recognition.
- To address the challenge of recognizing both seen and unseen SAR targets with limited training samples.
- To enable the interpretation of unseen categories without relying on simulation data.
Main Methods:
- Developed a novel OSR approach utilizing distribution construction and relation measures.
- Employed a graph convolutional network for distribution construction.
- The method operates without external simulation information, focusing on inherent data relationships.
Main Results:
- Achieved high recognition accuracy for seen SAR targets (above 95%).
- Demonstrated effective recognition of unseen SAR targets, reaching 67% in a three-class problem and 53% in a five-class problem.
- Showcased excellent interpretability for identifying unseen targets.
Conclusions:
- The proposed task-oriented OSR method is effective for SAR target recognition, even with limited training data and unseen categories.
- The approach provides a practical solution for real-world SAR applications where data collection is constrained.
- The method's ability to explain category similarity enhances the understanding of unrecognized targets.
