KDFF-DETR: Knowledge-Driven Feature Fusion DETR for Spaceborne Infrared Ship Detection
Spaceborne infrared imagery provides substantial data support for ship detection due to its adaptability and all-weather reconnaissance capabilities. However, in complex maritime environments, its inadequate semantic information and feature distinctiveness increase the risk of false alarms and misse...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11072708/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Spaceborne infrared imagery provides substantial data support for ship detection due to its adaptability and all-weather reconnaissance capabilities. However, in complex maritime environments, its inadequate semantic information and feature distinctiveness increase the risk of false alarms and missed detections. To address these challenges, we propose a knowledge-driven feature fusion detection transformer (KDFF-DETR) for spaceborne infrared ship detection. KDFF-DETR presents an enhancement of RT-DETR. Specifically, we design a hybrid Fourier contextual (HFC) encoder to model the adequate semantic representation of high-level features through frequency-domain features and masked supervision. We further develop a visual knowledge-supervised feature reconstruction (VKFR) module, which introduces feature operators to represent visual knowledge and reconstruct handcrafted features to enhance their distinctiveness. Besides, a contribution-balanced feature fusion (CBFF) method for query selection is proposed, dynamically adjusting the weight distribution of handcrafted and deep features during different training stages to improve feature fusion accuracy and generalization ability. Experimental results indicate that KDFF-DETR surpasses state-of-the-art (SOTA) methods in various public datasets. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |