On Bridge Surface Crack Detection Based on an Improved YOLO v3 Algorithm
Yuexin Zhang, Jie Huang, Fenghuang Cai
Abstract
An improved bridge surface crack detection algorithm based on a further developed You Only Look Once version 3 algorithm (YOLO v3) is proposed to realize the fast and accurate detection of bridge surface cracks for timely repair application scenarios. The proposed algorithm is combined with MobileNets and convolutional block attention module (CBAM), which can detect bridge surface cracks in real time. The standard convolution is replaced by the depthwise separable convolution of MobileNets so as to reduce the number of network parameters. Moreover, in order to solve the problem of precision decline caused by depthwise separable convolution, the inverted residual block of MobileNetV2 is introduced. Furthermore, the proposed algorithm selectively learn the feature by multiplying the attention map with the input feature map through CBAM, and focus on channel and spatial attention mechanisms simultaneously. Finally, the feasibility of the algorithm is verified by experiment.