PASSNet: Perception-Aware Semantic Strengthening Network for Infrared Small Target Detection

Infrared small target detection (IRSTD) is vital in military surveillance, aviation safety, and wildlife monitoring due to its strong anti-interference capability under complex backgrounds. However, the intrinsic properties of small targets, including weak semantic features, low contrast, and limited structural details, pose significant challenges for accurate detection. Existing deep learning-based methods suffer from unstable spatial perception in shallow layers, semantic dilution in deeper encoder stages, and a lack of high-level semantic guidance during decoding, which collectively lead to inaccurate localization and blurred boundaries. Moreover, most existing IRSTD methods fail to maintain a favorable balance between accuracy and efficiency under heterogeneous resource and frame-rate constraints. To address these issues, we propose PASSNet (Perception-Aware Semantic Strengthening Network), an efficient and effective IRSTD method that integrates enhancement mechanisms at three progressive levels-shallow spatial perception, deep layer channel semantics, and decoding reconstruction for real-world detection. Specifically, the Spatial Perception Enhancement Module (SPEM) leverages parallel multi-scale convolutions with spatial attention to stabilize shallow perception and better distinguish small targets from background clutter in shallow encoder layers. The Dynamic Semantic Aggregation Module (DSAM) employs graph-based inter-channel modeling to strengthen the deep semantic consistency of the encoder and mitigate feature dilution. Moreover, the Saliency-Guided Decoder Module (SGDM) performs nonlinear, saliency-aware channel reweighting to enhance structure reconstruction and reduce edge blurring. Extensive experiments on three widely used public datasets (NUDT-SIRST, NUAA-SIRST, and IRSTD-1k) demonstrate that the proposed PASSNet achieves superior performance in both detection accuracy and inference speed. In addition, the BSIRST_v2 dataset, built by our team, further validates PASSNet's promising generalization for multi-target detection in real-world scenarios.