An efficient bridge crack identification method based on a novel attention mechanism, transfer learning and lightweight network

In order to enhance the efficiency and accuracy of bridge crack detection, this paper proposes an efficient bridge crack identification method based on transfer learning, attention mechanism and lightweight convolutional neural network. First, based on the characteristics of small proportion of cracks and prominent edges in bridge images, a new attention mechanism CAM suitable for identifying bridge cracks is proposed and embedded into the lightweight network EfficientNetv2 to establish the CAM-EfficientNetv2 model. Second, for the problem of poor training effect and network overfitting due to the small sample dataset, the transfer learning method is introduced during network training. Finally, a bridge crack image dataset is constructed and the performance of the proposed method is analyzed in detail to verify the feasibility for crack identification. The experimental results demonstrate that the proposed method can achieve the overall precision, recall and F1-score of 96.46%, 96.59% and 96.64%, respectively. For different models constructed by other attention mechanisms embedded in EfficientNetv2 network, our method performs the best. Furthermore, it has significant advantages in terms of convergence speed and running speed, while also ensuring high accuracy. Our research outcome fully shows that this method can solve the problems of poor training performance, low detection accuracy, and slow detection speed caused by small sample datasets of existing detection networks in bridge crack recognition.