Infrared small target detection based on asymmetric convolution and adaptive loss

Le résumé fourni par la source

Infrared small target detection faces significant challenges including low signal-to-clutter ratio, complex background interference, and extreme foreground-background imbalance. Traditional detection methods often fail to capture subtle target features while maintaining robustness against environmental noise. YOLO-based detectors, though efficient, struggle with the unique spatial characteristics and scale sensitivity of infrared small targets. To bridge these gaps, we propose an enhanced YOLO framework that enhances robustness and computational efficiency through multi-stage optimization. First, an asymmetric convolution-based feature extraction network is designed, leveraging its spatial distribution characteristics to align with the Gaussian prior of targets, thereby enhancing feature representation while suppressing background clutter. Second, to tackle the extreme foreground-background imbalance, an adaptive threshold focal loss function is introduced, dynamically adjusting sample weights to improve the learning of sparse target features. Furthermore, a pyramid sparse Transformer module is incorporated, utilizing a hierarchical feature selection mechanism to optimize computational resource allocation. Experimental results demonstrate that the proposed method effectively reduces false alarms while maintaining high detection accuracy in complex scenarios, offering a practical solution for real-time infrared small target detection.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.