DFR-Net: Differential-Frequency Reliability Network for Infrared Small Target Detection

Abstract

Infrared small target detection (IRSTD) faces substantial challenges arising from extremely small target sizes, low contrast, and strong interference from complex backgrounds. Since small targets usually lack stable shape and texture information, their responses are easily weakened during successive downsampling. Meanwhile, low-level high-frequency features contain not only target details but also background edges, noisy bright spots, and local textures. Directly fusing these features may further amplify false-target responses during decoding. To address these issues, we propose a differential-frequency reliability network (DFR-Net) for IRSTD in complex scenes. DFR-Net adopts an encoder–decoder architecture and is designed around the preservation, enhancement, and selective restoration of target-related features. Specifically, the frequency-preserved adaptive downsampling module (FADM) integrates wavelet-decomposed features with pooled features during spatial resolution reduction to alleviate the loss of frequency cues associated with small targets. The differential-guided asymmetric feature module (DAFM) employs multiscale local differences to highlight response discrepancies between targets and their surrounding backgrounds, while incorporating directional context modeling to enhance target-related features. The reliability-guided Laplacian fusion module (RLFM) selects reliable details according to the consistency between high-level semantic information and low-level contextual information, thereby suppressing the propagation of unreliable high-frequency noise. Experimental results demonstrate that DFR-Net effectively improves the detection accuracy and robustness of infrared small targets in complex backgrounds.