Automatic Multiple Defect Detection Method for Underwater Concrete Bridge Piers based on Deep Learning

Timely detection of underwater structural defects is crucial for ensuring the stability of river-crossing bridges. Currently, underwater pier inspections are manual, labor-intensive, and subjective. Computer vision based on deep learning are considered advanced techniques for structural damage detection due to their end-to-end training process. This study propose a Swin-Res model fused with a self-attention residual network for the classification of underwater concrete defects in bridges. The Swin-Res model integrates a Swin-Transformer with self-attention mechanisms, a multi-layer residual structure (ResNet), and a multi-layer perceptron classification head. This model classifies defects such as exposed aggregate, mud attachment, exposed rebar, spalling, turbidity, and normal conditions in underwater bridge piers. The model was tested on the proposed image dataset of underwater bridge pier defects. Experimental results show that the Swin-Res model outperforms other comparative models on the test set in terms of accuracy, recall, and F1-score, achieving an average precision of 0.759, recall of 0.750, and accuracy of 0.750 for defect detection. Furthermore, comparative results with various existing classification models indicate that the method surpasses these models. This study provides an efficient and reliable approach for the intelligent detection and classification of challenging underwater bridge defects.