GLAF-DETR: Detection Transformer With Global-Local Adaptive Fusion Attention for Infrared Maritime Object Detection
- 11 August 2025
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. PP, 1
Abstract
Infrared maritime object detection is a crucial technology for sea surface monitoring in low-light conditions within maritime Internet of Things (IoT) systems. In practical applications, this task faces significant challenges, including diverse target sizes and stringent real-time processing requirements. To address these challenges, a DEtection TRansformer with Global-Local Adaptive Fusion attention for infrared maritime object detection (GLAF-DETR) is proposed. The Global-Local Adaptive Fusion (GLAF) attention mechanism is designed to capture both global contextual information and fine local details of objects. GLAF dynamically adjusts attention across regions by integrating long-range dependencies with short-range positional information, significantly enhancing detection performance for targets of varying sizes in complex maritime environments. In addition, the Dynamic Adaptive Multiscale Feature Fusion (DAMFF) module is proposed to promote cross-channel interaction among multiscale features. Guided by GLAF, DAMFF dynamically fuses these features, further enhancing the accuracy of multiscale object detection. The lightweight HGNetv2-IRLight backbone is designed to minimize network complexity and ensure real-time performance by reducing redundant information while maintaining strong infrared feature extraction. Extensive experiments conducted on an infrared maritime object dataset show that GLAF-DETR surpasses state-of-the-art methods in both detection accuracy and inference speed. It demonstrates outstanding performance, particularly in detecting objects across different scales, offering enhanced accuracy and robustness in challenging maritime scenarios.Keywords
Funding Information
- National Natural Science Foundation of China (62463004, U2141234, U24A20260)
- Hainan Province Science and Technology Special Fund (ZDYF2024GXJS003)
- National Science and Technology Major Project (2022ZD0119901)
This publication has 25 references indexed in Scilit:
- YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information2024
- Coordinate Attention for Efficient Mobile Network Design2021
- End-to-End Object Detection with Transformers2020
- ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks2020
- EfficientDet: Scalable and Efficient Object Detection2020
- Squeeze-and-Excitation Networks2018
- Path Aggregation Network for Instance Segmentation2018
- Feature Pyramid Networks for Object Detection2017
- You Only Look Once: Unified, Real-Time Object Detection2016
- Microsoft COCO: Common Objects in Context2014