Frequency-domain and spatial-perception collaborative learning for infrared small target detection

Abstract

Infrared small-target detection is widely used in early warning and remote monitoring. The targets are usually extremely weak, extremely small, and easily obscured by the chaotic background. However, existing methods still have poor detection accuracy in low-contrast and highly interfering scenarios, and many methods rely mainly on spatial-domain features while ignoring useful frequency-domain clues. Moreover, the common multi-scale fusion based on simple concatenation or addition cannot fully leverage the complementary strengths of different layers and domains, leading to feature interference and missed detections. To solve these problems, we propose an efficient detection network called the Frequency-Domain and Spatial-Perception Collaborative Learning Network (FSCLNet). This is a spatial-frequency dual-domain collaborative network with improved multi-scale fusion, consisting of the following three modules: Dual-domain Guided Feature Extraction Module (DGFEM), which jointly learns spatial structure and frequency features to enhance the saliency of the target; Joint Domain Perception Cascade Module (JDPCM), which refines features through cascaded cross-domain interactions to improve robustness; and Efficient Feature Fusion Module (EFFU), which performs more stable and selective feature fusion between different scales. Experiments across two datasets show that our method achieves the best overall performance compared to state-of-the-art methods.