PHFNet: A Parallel Hybrid Fusion Network With Linear Attention for Infrared Small Target Detection

Infrared small target detection (IRSTD) is challenging due to weak targets and complex backgrounds, which undermine the performance of purely convolutional neural networks (CNNs). While recent hybrid CNN–Transformer architectures attempt to integrate local details with long-range context, their global perception tends to be incomplete and unstable, primarily owing to the prohibitive computational cost of full global modeling. To address this limitation, we propose PHFNet, a parallel hybrid fusion network for IRSTD that enables continuous local–global feature collaboration across all network stages. Specifically, PHFNet incorporates a CNN branch dedicated to extracting fine local details and preserving target structures, alongside a Transformer branch equipped with a novel Local-Compensated Linear Attention (LCLA). LCLA introduces adaptive local compensation into linear attention to preserve target-related local cues during global aggregation and suppress structured local interference, while structured spatial interaction further reduces the representational gap across the two branches. In addition, a Bidirectional Multi-scale Refinement Attention (BMRA) module is introduced to refine fused features through multi-scale contextual aggregation and spatial constraints, thereby enhancing representation stability and suppressing structured background interference. Extensive experiments on multiple public datasets demonstrate that the proposed PHFNet consistently outperforms existing IRSTD methods. The code will be released at https://github.com/Llt201/PHFNet.