Automatic crack recognition for concrete bridges using a fully convolutional neural network and naive Bayes data fusion based on a visual detection system
Gang Li,Qiangwei Liu,Shanmeng Zhao,Wenting Qiao,Xueli Ren
DOI: https://doi.org/10.1088/1361-6501/ab79c8
IF: 2.398
2020-05-05
Measurement Science and Technology
Abstract:Regular inspections of bridge substructures are very important for evaluating bridge health, since early detection and assessment offer the best chances of bridge repair. However, the traditional inspection methods of checking defects with visual features cannot meet engineering needs sufficiently. Although deep-learning methods have recently demonstrated a remarkable improvement in image classification and recognition, there are still difficulties, such as the countless parameters and large model training sets needed by these methods. In this paper, we propose a novel crack extraction algorithm for automatic segmentation of cracks and noise using multi-layer features extracted from a fully convolutional network and a naive Bayes data fusion (NB-FCN) model. The bridge images in both the training and testing datasets are taken using an in-house designed high-precision image acquisition device, called Bridge Substructure Detection 10 (BSD-10). BSD-10 is applied to collect 7200 images from ten existing bridges under different illuminants and distances. After gathering the crack datasets, the crack and noise models of the NB-FCN are trained, respectively, with multiple iterations. Next, the skeleton and continuous boundary of a crack are recognized. Then the crack length and width are calculated using electronic distance measurement to verify the error rate of the proposed method. Compared to up-to-date machine-learning-based algorithms, i.e. the crack tree algorithm, the random structured forests algorithm, the relatively competitive convolutional neural networks algorithm, and the fusion convolutional neural network algorithm, the significant superiority of the NB-FCN algorithm in terms of recognition accuracy, computation time, and error rates is illustrated based on different types of crack images of handwriting, peel off, water stains and repair traces. The NB-FCN algorithm is verified with 7200 datasets of bridge substructures collected from 20 in-service bridges under various circumstances. In general, the recognition results show that the proposed algorithm demonstrates a remarkable performance compared to other recent algorithms.
engineering, multidisciplinary,instruments & instrumentation