QGCANet: Query-guided global context aggregation network for infrared small target detection

Abstract

Infrared small target detection faces challenges: targets are easily overwhelmed by background clutter in low SNR environments, have insufficient spatial representation due to tiny size, and suffer boundary ambiguity from lacking global context. Existing methods struggle to model long-range dependencies between targets and clutter, leading to inaccurate detection. To address this, we propose the Query-Guided Global Context Aggregation Network (QGCANet). It features a Query-Enhanced Global Feature Extraction Module that employs Transformer-inspired query vectors to guide multi-layer contextual feature aggregation, leveraging long-range dependencies to distinguish targets. To resolve the low-contrast problem induced by the low SNR of infrared small targets, we propose an Adaptive Contrast Sensitivity Enhancement Module, which amplifies the distinction between targets and backgrounds through dynamic local contrast enhancement, making faint targets more prominent, while utilizing query-guided spatial attention weights to suppress noise interference. Furthermore, to mitigate insufficient representation capacity caused by semantic conflicts across different stages, we propose a Query-Guided Multi-scale Feature Fusion module, which reduces the semantic gap between encoder and decoder stages through multi-scale feature extraction and query-based dynamic weighting. Extensive experiments demonstrate QGCANet significantly outperforms state-of-the-art methods.