FRF-YOLOv11: frequency-domain robust feature fusion for low-light object detection based on YOLOv11

Abstract

Low-light object detection remains challenging due to severe illumination degradation, non-uniform local lighting, and amplified noise. Conventional low-light enhancement methods mainly operate in the spatial domain and often improve visual brightness without consistently benefiting downstream detection. In this paper, we propose a frequency-domain robust feature fusion framework for low-light object detection based on YOLOv11. A Frequency-domain Robust Branch (FRB) is introduced to extract illumination-invariant representations from log-transformed RGB channels through frequency-domain channel differencing, zero-DC suppression, and learnable radial-basis filtering. The resulting frequency features are adaptively fused with shallow spatial features and then fed into YOLOv11 for end-to-end detection. The proposed design is lightweight, detector-oriented, and easy to integrate into existing one-stage detectors. Experiments on the ExDark dataset show that the proposed method improves mAP@0.5 by 5.5% points and mAP@0.5:0.95 by 6.1% points over the YOLOv11 baseline, while maintaining efficient inference. Ablation studies further verify the effectiveness of the frequency branch and the adaptive fusion module.