A vision-based method for crack detection in gusset plate welded joints of steel bridges using deep convolutional neural networks

There has been a growing demand for early detection of fatigue cracks in gusset plate joints in steel bridges. The present study develops a robust method for crack detection using the concept of transfer learning as an alternative to training an original neural network. Three standard deep learning methods of training a crack classifier using 1) a shallow convolutional neural network built from scratch, 2) the output features of the VGG16 network architecture previously trained on the general ImageNet dataset, and 3) the fine-tuned top layer of VGG16 are investigated. Data augmentation is used to reduce overfitting caused by the limited and imbalanced training dataset. The image dataset includes both fatigue test photographs and actual inspection photographs captured under uncontrolled distance, lighting, angle, and blurriness conditions. Data augmentation increased the accuracy index by 5%, 2%, and 5%, respectively, for the three methods. The fine-tuning methods achieved the best precision and recall scores and the best robustness. Slightly fine-tuning a well-trained fully connected layer together with the top convolutional layer of the VGG16 model in combination with data augmentation achieved the best performance for crack detection in gusset plate joints of steel bridges. Transfer learning methods using the output features of VGG16 achieved both improved accuracy and robustness and, therefore, are recommended for training a crack classifier.