A lightweight infrared remote sensing architecture for enhanced small target detection using improved DETR with CST...

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Infrared remote sensing (IRS) ship detection is challenged by low resolution and noise, especially for small targets.
• Existing methods struggle with inter-channel imbalance and feature loss in infrared images.
• Small targets are often obscured by background noise, hindering accurate detection.

Purpose of the Study:

• To propose a novel lightweight architecture, RT-DETR-CST, for enhanced small target detection in IRS.
• To address challenges of inter-channel information imbalance and feature degradation in infrared ship detection.
• To improve accuracy, reduce model size, and increase inference speed for real-time applications.

Main Methods:

• Developed a Cross-Channel Feature Attention Network (CFAN) for weighted feature fusion and noise suppression.
• Introduced a Scale-Wise Feature Network (SWN) using depthwise separable convolutions and stochastic depth for multi-scale feature extraction.
• Integrated a Texture/Detail Capture Network (TCN) for edge and detail preservation through linear decomposition and channel fusion.

Main Results:

• RT-DETR-CST achieved an mAP0.5 of 89.4% on ISDD datasets, a 4.9% improvement over RT-DETR.
• Model size was reduced by 41.5% to 23.7 MB, with an inference speed of 207.2 FPS.
• Ablation studies confirmed the effectiveness of each module, demonstrating superior accuracy, lightweight design, and real-time performance.

Conclusions:

• RT-DETR-CST offers a highly effective solution for small target detection in infrared ship remote sensing.
• The proposed architecture demonstrates excellent generalization capabilities across different datasets and sensing modalities (IRS and SAR).
• The model achieves a strong balance between accuracy, efficiency, and robustness for real-time remote sensing applications.