FDBGF-Net: Frequency-Decoupled Bidirectional Gated Fusion for Infrared–Visible Object Detection Under Degradation in UAV Imagery
Abstract
Infrared–visible fusion is increasingly used to improve object detection robustness in uncrewed aerial vehicle (UAV) remote sensing. However, real-world UAV imagery is often affected by co-occurring degradations, such as blur, rain, snow, fog, strong illumination, and electromagnetic interference. These disturbances may simultaneously corrupt both modalities, weaken cross-modal complementarity, and make conventional fusion schemes unreliable. To address this problem, we propose FDBGF-Net, a frequency-decoupled bidirectional gated fusion network for infrared–visible object detection under compound degradation. The proposed framework introduces three coordinated components. First, a dual-modal gated fusion network performs P3-level cross-modal alignment through selective state-space modeling, preserving geometric details before high-level semantic compression. Second, a degradation-aware gated fusion network employs a bidirectional gated feature pyramid and a spatial adaptive router to dynamically aggregate multiscale features according to modality reliability. Third, a Wavelet-Decoupled Frequency Attention module separates fused representations into low-frequency structural and high-frequency detail components, and adaptively recalibrates their contributions using dual-path channel gating. We also construct a benchmark containing 3711 infrared–visible image pairs, five target categories, six degradation types, and three severity levels. Experimental results show that FDBGF-Net achieves 79.8% mAP<inline-formula><tex-math notation="LaTeX">$_{50}$</tex-math></inline-formula> on the proposed benchmark, outperforming DPAL-R by 1.5 percentage points at 76.3 FPS. It also reaches 79.6% mAP<inline-formula><tex-math notation="LaTeX">$_{50}$</tex-math></inline-formula> on VEDAI and 88.1% mAP<inline-formula><tex-math notation="LaTeX">$_{50}$</tex-math></inline-formula> on aligned FLIR after independent target-dataset retraining, demonstrating consistent cross-dataset applicability across substantially different data distributions.