基于深度学习的钢桥病害分割与量化

Segmentation and quantification of steel bridge defects based on deep learning

为解决传统钢桥表面病害难以量化的难题,提出了一种基于深度学习的钢桥病害分割与量化方法.该方法以SOLOv2实例分割网络为基础,将其原ResNet骨干网络替换为VoVNet57网络进行检测与分割.依据分割的病害掩码,增添检测框,计算病害像素值,并转化为实际面积,完成量化.采集涂层劣化、腐蚀、焊缝开裂等钢桥病害图像共计314幅,数据增强后,得到3241幅病害图像,按4∶1划分训练集和测试集,用Labelme软件进行人工标注.经20000次训练后,得到63.45%的平均精度(mAP),而TensorMask、BlendMask、未改进SOLOv2的训练mAP值分别为55.89%、56.67%、58.25%.同时,将网络输出的病害尺寸量化值和病害尺寸实际值进行对比得到,涂层劣化和腐蚀病害面积相对误差集中在10%以内,焊缝开裂长度相对误差集中在8%以内.所提出的方法可实现钢桥多病害快速、精确的分割与量化,为钢桥智能检测评估提供了技术支撑.

To solve the problem that it is difficult to quantify the surface defects of traditional steel bridges,a segmentation and quantification method of steel bridge defects based on deep learning is proposed.Based on the Solov2 instance segmentation network,the original ResNet backbone network is replaced by VoVNet57 network for detection and segmentation.According to the segmented disease mask,a detection frame is added,the defect pixel value is calculated and transformed into the actual area to complete the quantification.A total of 314 images of steel bridge defects such as coating deterioration,corrosion and weld cracking were collected.After data enhancement,3241 defect images were obtained.The training set and test set were divided according to 4∶1,and manually marked with Labelme software.After 20000 training sessions,the mean average accuracy(mAP)of 63.45%is obtained,while the training mAP values of TensorMask,BlendMask and unmodified Solov2 are 55.89%,56.67%and 58.25%,respectively.By comparing the quantitative value of the defect size output by the network with the actual value of the defect size,the relative error of the coating deterioration and corrosion defect area is concentrated within 10%,and the relative error of the weld crack length is concentrated within 8%.Results show that the proposed method can be used to realize the rapid and accurate segmentation and quantification of multiple defects of steel bridges,and provides technical support for the intelligent detection and evaluation of steel bridges.