Lightweight Bridge Crack Detection with YOLO-STCDE

Bridge crack detection is a critical task in structural health monitoring, yet existing deep learning methods often suffer from parameter redundancy in backbone networks, insufficient adaptivity to complex background interference in feature enhancement modules, and feature conflicts between classification and regression subtasks in coupled detection heads. To address these challenges, this paper proposes YOLO-STCDE, an improved lightweight detection model built upon YOLOv12. The model integrates three synergistic architectural innovations: (1) ST-Net, a lightweight backbone that replaces the original R-ELAN with a star operation-based design incorporating DynamicTanh activation for implicit high-dimensional feature mapping and adaptive amplitude calibration; (2) A2C2f-CD, a dual-dynamic gated neck module that embeds DynamicTanh and Convolutional Gated Linear Units into the A2C2f architecture, enhancing crack feature discrimination under complex backgrounds; and (3) Efficient-Detect, a decoupled detection head with a shared convolution stem that eliminates task conflict while substantially compressing parameter overhead. Extensive experiments on the bridge crack dataset demonstrate that YOLO-STCDE achieves 91.6% mAP@0.5 and 69.4% mAP@0.5:0.95 with only 2.25 M parameters and 5.6 GFLOPs, representing improvements of 3.2 and 9.5 percentage points over the YOLOv12n baseline, respectively, while simultaneously reducing the parameter count by 10.4%. Compared with state-of-the-art lightweight detectors including YOLOv11n and Hyper-YOLO, YOLO-STCDE attains the highest detection accuracy with the smallest model footprint, demonstrating an optimal balance between accuracy and efficiency for real-world bridge crack inspection.