CSFPR-RTDETR-CR: A Causal Intervention Enhanced Framework for Infrared UAV Small Target Detection with Feature...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Infrared UAV small target detection is vital for reconnaissance and monitoring.
• Challenges include small target size, weak texture, and complex backgrounds, leading to model bias and poor performance.

Purpose of the Study:

• To develop an enhanced detection framework using causal reasoning to overcome feature bias in infrared small target detection.
• To improve the generalization and accuracy of deep learning models in complex environments.

Main Methods:

• Proposed an enhanced detection framework building on the CSFPR-RTDETR detector, incorporating causal reasoning principles.
• Implemented a three-path feature debiasing approach: causal data augmentation, counterfactual reasoning module, and causal attention mechanism.
• Utilized frequency perturbations and counterfactual samples to separate causal and non-causal features.

Main Results:

• Achieved a 5.6% improvement in mAP@50 and a 1.8% improvement in mAP@50:95 on the HIT-UAV dataset.
• Demonstrated enhanced feature discrimination and overall detection performance through visualization analysis.
• Reduced spurious correlations between targets and backgrounds, leading to fewer false positives and missed detections.

Conclusions:

• The causal reasoning framework effectively debiases features, enhancing robustness and accuracy in infrared UAV small target detection.
• The proposed methods significantly improve detection performance in challenging scenarios with complex backgrounds.
• This approach offers a promising direction for developing more reliable and generalizable object detection systems.