FADet: A Frequency-Aware Detection Framework for Infrared Small Target Detection

Abstract

Infrared small target detection (IRSTD) remains a critical yet highly challenging task in the field of object detection. Due to the tiny target size and the absence of rich texture information, general-purpose detectors often suffer substantial performance degradation when applied to this task. This performance degradation is mainly due to their limited ability to extract discriminative features, resulting in frequent missed detections and false alarms that compromise the reliability of detection systems. To address these challenges, we propose FADet, a novel Frequency-Aware Detection framework specifically designed to capture the unique representational characteristics of small targets. Specifically, we introduce a Frequency-Guided Visual Encoder that leverages the Haar Wavelet Transform to explicitly decompose spatial features into high- and low-frequency components. An attention mask is then derived from the high-frequency components to selectively preserve informative fine-grained details. This process effectively alleviates the over-smoothing effect typically induced by convolutional operations, thereby significantly enhancing the saliency of small targets in the detection framework. Furthermore, we propose a Multiscale Feature Gather-Distribute module that aggregates multiscale semantic cues and redistributes them across different feature hierarchies, thereby enabling more effective feature interaction and fusion. Extensive experiments on three public benchmark datasets (e.g., NUAA-SIRST, NUDT-SIRST, and IRSTD-1 K) demonstrate that FADet achieves superior performance, setting new state-of-the-art results in infrared small target detection.