Lightweight concrete bridge damage detection by improved YOLOv5 and channel pruning algorithm
Hao Li, Jianxi Yang, Shixin Jiang, Xiaoxia Yang · Measurement Science and Technology · 2025
Abstract In the field of automated concrete damage detection, deep learning methods face challenges in real-time performance and accuracy. This paper proposes a lightweight method for concrete bridge damage detection, combining an enhanced YOLOv5 model with channel pruning algorithms. Firstly, ShuffleNetv2 is used as the feature extraction backbone, reducing network parameters. Secondly, bi-directional feature pyramid network fusion improves accuracy for small object damage by merging feature maps with contextual information. The model is then trained with channel sparse regularization and pruned based on sparsity rates. Finally, fine-tuning identifies optimal pruning parameters. Evaluated on a dataset of 2468 high-definition images containing four types of concrete bridge damage, the model achieves 72.92% precision, 77.85% recall, and 78.75% mean average precision (mAP), with the pruned model achieving 76.14% mAP. The model size is reduced from 10.26 M to 4.24 M, meeting real-time, high-accuracy requirements with strong generalizability for bridge detection.