LAUSR

Abstract Manual visual inspection is customarily used to identify and evaluate service status of bridges. However, current procedures followed by human inspectors demand long inspection time to access bridges. Also, highly relying on subjective or empirical knowledge of inspectors may induce false evaluation. To overcome these challenges, a vision-based method is presented for bridge defects detection using transfer learning and convolutional neural networks (CNNs), which could automatically analyze and identify a large volume of collected images. Firstly, typical defect images are preprocessed by means of image processing techniques (IPTs). Secondly, the transfer learning model is trained on 1180 images with arbitrary sizes and pixel resolutions. Thirdly, the trained model is tested on 134 images taken from different bridges which are not used in training and validation sets, and finally recorded the accuracy of 97.8% for testing set. Comparative studies are conducted to examine the performance of the proposed approach using classical machine learning algorithms (MLAs) and features extracted with existing hand-craft methods. The results demonstrate that the proposed approach shows quite better performance in accuracy and efficiency.