基于卷积神经网络的桥梁裂缝检测方法

Bridge crack detection method based on convolution neural network

针对桥梁裂缝固有特征及检测过程的局限性,引入基于卷积神经网络的YOLOv3单阶段目标检测算法,并对YOLOV3网络的多尺度预测模块进行改进,充分利用浅层特征,提升小裂缝检测精度。通过聚类算法对数据集进行聚类,得到适用于桥梁裂缝特征的先验框尺寸。数据集方面引入生成对抗网络对桥梁裂缝数据集进行扩增。实验结果表明,在相同数据集和迭代次数下,改进YOLOv3网络裂缝检测精度可达0.9302,比原YOLOv3提高0.0137。

In view of the inherent characteristics of bridge cracks and the limitations of the detection process,a single-stage target detection algorithm based on convolution neural network was introduced,and the multi-scale prediction module of YOLOv3 network was improved to make full use of the shallow features and improve the detection accuracy of small cracks.Through clustering algorithm,the data set was clustered to get the prior frame size which was suitable for the characteristics of bridge cracks.On the aspect of data set,anti generation network was introduced to expand the data set of bridge cracks.Experimental results show that under the same data set and iteration times,the crack detection accuracy of the improved YOLOv3 network reaches 0.9302,which is 0.0137 higher than that of the original YOLOv3 network.