Research on Few-Shot Class-Incremental Recognition Technology for SAR Images

Click the blue “Radar Journal” to follow

From October 22 to October 30, 2023, the 3rd Radar Journal Doctoral Forum was successfully held! Doctoral student Zhao Yan from National University of Defense Technology presented an academic report titled “Research on Few-Shot Class-Incremental Recognition Technology for SAR Images” at the SAR Image Interpretation Technology Sub-Forum on the morning of October 28.

Report Overview

Currently, deep learning technology has shown exceptional advantages in the field of SAR image target recognition. However, most of these algorithms are designed for ideal closed recognition scenarios and lack the ability to continuously and agilely learn and update knowledge of new target categories in open and dynamic real-world environments. Furthermore, due to observational constraints, the small sample characteristics of targets are significant in most non-cooperative scenarios. Therefore, researching Few-Shot Class-Incremental Learning (FSCIL) SAR image target recognition technology for real-world application scenarios has significant research significance and application value.

This study is the first to explore the application of FSCIL in the field of SAR target recognition, revealing domain-specific challenges and proposing a SAR image target FSCIL framework based on Cosine Prototype Learning (CPL). CPL relies on the accurate perception and intrinsic correlation of structurally scattered information sensitive to the target’s azimuth angle in a cosine semantic space, and designs relevant loss functions and learning strategies to balance the algorithm’s “stability” in representing old knowledge and “plasticity” in learning new knowledge. The algorithm demonstrates significant advantages over advanced benchmark algorithms in various FSCIL scenarios constructed from the MSTAR dataset, validating its effectiveness in addressing the FSCIL problem in SAR ATR.

Report Video

Video playback address on computer (Radar Journal Bilibili Learning Platform):

3rd Radar Journal Doctoral Forum | Research on Few-Shot Class-Incremental Recognition Technology for SAR Images

Report PPT

This report PPT has a total of 25 slides.

Doctoral Introduction

Zhao Yan, a doctoral student at the National University of Defense Technology, supervised by Professor Kuang Gangyao. His research directions include SAR image target detection and recognition, SAR image target incremental learning, few-shot learning, etc.

The relevant section of the website “Academic Reports” can be accessed by clicking “Read the original text“.

↓↓↓

The 3rd Radar Journal Doctoral Forum was successfully held!

[Latest Achievement] Visualization Method for CNN Model Targeting SAR Image Classification (Video)

[Latest Review] Research Progress and Prospects of Intelligent Detection and Recognition Technology for Aircraft Targets in SAR Images (Video)

[Achievement] FCOSR: An Anchor-Free SAR Image Arbitrary Orientation Ship Target Detection Network (Video)

[Achievement] SAR Image Ship Target Autofocus Algorithm Based on Adaptive Momentum Estimation Optimizer and Spatial Minimum Entropy Criterion

[Achievement] Multi-Feature Fusion Method for Sea Surface Target Detection Based on Deep Learning (Video)

Future Star | Ship Targets in PolSAR Images Based on Differences in Polarization Characteristics (Video)

Doctoral Forum | Preliminary Exploration of Target Detection and Recognition Methods for Multi-Source Spaceborne SAR Images (Video)

Doctoral Forum | SAR Image Target Recognition Technology Combining Electromagnetic Scattering Features and Deep Features (Video)