DS-YOLO: A dual-domain synergistic YOLO for infrared–visible object detection
Chunhua Zhu, Zhihua Liu, Yongfang Li, Ning Li · Infrared Physics & Technology · 2026
Infrared–visible object detection improves perception in low-illumination and adverse environments by combining the robustness of infrared imagery with the detail richness of visible imagery. However, existing methods mainly rely on spatial-domain fusion and neglect spectral characteristics, leading to insufficient feature decoupling and cross-scale inconsistency in low-contrast and cluttered scenes, which degrades small-target detection and localization. To address this issue, we propose a Dual-Domain Synergistic YOLO framework (DS-YOLO) that jointly models frequency-domain spectral responses and spatial-domain structure for more effective cross-modal representation. A Frequency Collaborative Alignment Module (FCAM) transforms low-level features into the frequency domain via fast Fourier transform, explicitly decoupling amplitude and phase components. Amplitude-guided energy weighting compensates for attenuated high-frequency responses in infrared imagery, while phase-aware refinement enhances feature-level structural consistency between modalities. To produce high-level semantic representations, the Spatial Expansion Fusion Module (SEFM) adopts a multi-branch design incorporating dilated convolutions, enabling the capture of large-scale thermal patterns and their contextual dependencies. A Feature Pyramid Interaction Network further enhances bidirectional cross-scale information propagation. By improving cross-modal complementarity, DS-YOLO achieves improved detection accuracy and robustness under complex conditions, with mAP50 scores of 96.5%, 85.8%, and 83.1% on the LLVIP, M3FD, and FLIR datasets, respectively.