Asymmetric visible-infrared cross-fusion for robust bolt localization and missing detection

Accurate and reliable detection of missing bolts remains challenging under exposure anomalies, low illumination, and cluttered backgrounds. Visible images are sensitive to illumination changes and may lose texture and detail, whereas infrared images are more illumination-robust but lack the detail and resolution required for accurate standalone detection. Existing general multimodal fusion models are primarily designed for universal detection scenarios, which fail to adapt to the unique feature distribution characteristics of bolt targets with weak textures and easily degraded deep features. Such methods cannot effectively extract critical fine-grained target features, leading to prominent performance limitations in industrial bolt detection tasks. To address these complementary limitations of dual-modal data and the drawbacks of existing fusion methods, we propose Dual-modal Asymmetric Cross-Fusion YOLO (DACF-YOLO), a lightweight network for bolt localization and missing detection. Modality-specific asymmetric branches and bidirectional shallow high-frequency feature injection exploit cross-modal complementarity and compensate for information loss. A Shallow Feature Screening and Enhancement module suppresses redundant texture noise and background interference, improving early feature quality. A Cross-Modal Guided Fusion mechanism dynamically assigns weights according to modality-specific feature quality, enabling adaptive complementary enhancement. A parameter-shared Lightweight Detail-Enhanced Detection Head preserves critical bolt textures and contours while reducing model complexity. Experiments on a registered visible-infrared bolt dataset show that DACF-YOLO achieves 95.3% mAP50 and 63.2% mAP50-95 with 3.28 M parameters and 9.7 GFLOPs, demonstrating a favorable balance between detection accuracy and computational cost under challenging industrial conditions.