MCL-YOLO: A Multi-Module Collaborative Lightweight Object Detection Method for Bridge Crack Detection
TL;DR
A multi-module collaborative lightweight model (MCL-YOLO) based on YOLOv12 is proposed to reduce computational complexity while preserving critical information during feature downsampling to enhance the representation of slender, curved, and branched crack patterns.
Abstract
Bridge surface cracks are important early indicators of structural performance degradation. However, affected by complex environmental interferences and irregular morphologies, existing models still fall short in micro-crack recognition, accurate bounding-box localization, and lightweight. To address these challenges, this study proposes a multi-module collaborative lightweight model (MCL-YOLO) based on YOLOv12. Specifically, the existing ADown module from YOLOv9 is incorporated into the YOLOv12 architecture to reduce computational complexity while preserving critical information during feature downsampling. To enhance the representation of slender, curved, and branched crack patterns, a C3k2-RFAConv module is designed by integrating a receptive-field attention mechanism. Furthermore, an iEMA module is embedded before the high-resolution detection branch to strengthen the semantic response to weak-texture cracks. A bridge crack dataset containing 4029 images was constructed to evaluate the proposed model. Experimental results show that MCL-YOLO achieves Precision, Recall, mAP@50, and mAP@50:95 values of 0.891, 0.757, 0.844, and 0.675, with 5.5 GFLOPs, 2.240 M parameters, and a model-file size of 4.689 M. Compared with the YOLOv12n baseline, MCL-YOLO improves the four detection metrics by 2.30%, 4.56%, 4.07%, and 3.21%, while reducing GFLOPs and parameter count (Params) by 12.70% and 12.77%, respectively. Ablation experiments, model version comparisons, attention mechanism comparisons, and qualitative detection results collectively verify the effectiveness of the integrated architectural modifications.