Dual-Pathway Feature Separation and Gated Fusion Network for Infrared Small Target Detection

Infrared small target detection (IRSTD) plays a vital role in infrared search and tracking (IRST), enabling intelligent systems to accurately detect dim and small targets within cluttered thermal environments. However, most existing deep learning approaches for IRSTD employ a unified-pathway architecture that conflates saliency and edge information within a shared representation space. This limitation causes feature entanglement, hindering the network’s capacity to accurately separate and represent global saliency and fine-grained edge contours. To overcome these challenges, we propose LoveNet, a dual-pathway network architecture that explicitly separates feature learning into two specialized branches. The first is a multiscale saliency learning branch designed to extract comprehensive structural and contrast information, capturing the global context of targets. The second is a fixed-scale edge learning branch aimed at preserving spatial details and enhancing the precision of edge contour delineation. To integrate the heterogeneous features extracted by two branches, a gated feature fusion mechanism is proposed to adaptively combine saliency and edge representations based on their spatial and semantic relevance. Furthermore, to provide robust and comprehensive supervision, a hybrid supervision strategy (HSS) is designed to guide the learning process of hierarchical feature representations. Experiments on the NUDT-SIRST, IRSTD-1k, and SIRST datasets demonstrate that LoveNet consistently achieves the best segmentation performance compared to the state-of-the-art methods, while maintaining a lightweight structure suitable for real-time applications.