Enhancing Target and Rectifying Clutter: A Structure-Guided Frequency Rectification Network for Infrared Small-Target Detection

Abstract

Infrared small-target detection (IRSTD) remains challenging under low signal-to-noise ratio conditions, where weak target responses are easily submerged by complex background clutter. Existing enhancement-oriented methods may inadvertently activate target-like clutter in complex backgrounds. Once propagated through feature aggregation, these mixed activations can degrade mask reconstruction and increase false alarms. In multiscale high-frequency representations, target and clutter regions may both produce pronounced responses, but their dominant response locations exhibit distinct behaviors across neighboring scales. Target-related peaks tend to remain close to their reference-scale locations, whereas clutter-related peaks exhibit larger cross-scale displacements. This distinction suggests that weak target enhancement and clutter response rectification should be considered jointly. Accordingly, we propose a structure-guided frequency rectification network (SGFR-Net) for IRSTD. Specifically, a progressive contextual differential enhancement block is designed to strengthen weak local target responses and reduce target attenuation during feature propagation. Since the enhanced responses may still contain target-like clutter, an Edge-Guided Frequency Interaction Refinement module is introduced to refine clutter-contaminated high-frequency representations through edge-guided compact cue selection and cross-scale high-frequency interaction. A Cross-Level Feature Interaction Fusion module further performs selective multilevel aggregation to reduce feature inconsistency during mask reconstruction. Extensive experiments on SIRST, NUDT-SIRST, and IRSTD-1 K demonstrate that SGFR-Net achieves strong segmentation accuracy and robust detection performance under complex backgrounds.