基于特征金字塔网络的超大尺寸图像裂缝识别检测方法

来源:用户上传      作者:舒江鹏 李俊 马亥波 段元锋 赵唯坚

　　摘 要：基于D像分析的裂缝自动检测识别一直是桥梁结构健康检测的热点问题之一。深度学习作为裂缝检测的重要解决方法，需要大量数据支持。公开数据集提供的小尺寸裂缝图像不足以解决超大尺寸细长裂缝图像的检测问题。提出一个基于特征金字塔深度学习网络的超大尺寸图像中细长裂缝的检测方法。通过对编码器提取的4个不同层次的特征图分别进行预测，网络能够实现对细小裂缝的高精度分割。试验使用120张大小为3 264×4 928像素的桥钢箱梁表面裂缝图像对特征金字塔网络进行训练、测试;并将获得的训练模型与通过双线性插值方法缩放图像至1 600×2 400像素和2 112×3 168像素两种规格生成的数据集训练后的模型进行对比。结果表明：该方法在对比测试中能够获得最高的裂缝检测交并比（IoU）为0.78，最低的Dice Loss为0.12。测试中，裂缝检测图像显示，缩放图像会导致部分裂缝信息的丢失，该方法能稳定地保留裂缝信息，并实现复杂背景下超大尺寸图像中细长裂缝的高精度自动检测。
　　关键词：裂缝检测;深度学习;超大尺寸图像;特征金字塔网络
　　中图分类号：U446.2 文献标志码：A 文章编号：2096-6717（2022）03-0029-08
　　收稿日期：2021-04-28
　　基金项目：国家自然科学基金（U1709216）;国家重点研发计划（2018YFE0125400）
　　作者简介：舒江鹏（1987- ），男，博士，研究员，主要从事结构智能检测和评估，E-mail：jpeshu@zju.edu.cn。
　　Received：2021-04-28
　　Foundation items：National Natural Science Foundation of China （No. U1709216）; The National Key Research and Development Program of China （No. 2018YFE0125400）
　　Author brief：SHU Jiangpeng（1987- ）， PhD， professor， main research interests： inspection and assessment of structures， E-mail： jpeshu@zju.edu.cn.
　　Crack detection method based on feature pyramid network for super large-scale images
　　SHU Jiangpenga，b， LI Juna，c， MA Haiboa， DUAN Yuanfenga， ZHAO Weijiana
　　（a. College of Civil Engineering and Architecture; b. Center for Balance Architecture; c. The Architectural Design & Research Institute of Zhejiang University Co.， Ltd.， Zhejiang University， Hangzhou 310058， P. R. China）
　　Abstract： Automatic crack detection based on image analysis is a hot issue in bridge structure health inspection. Crack segmentation based on deep learning is a significant solution， which needs lots of database. The small-scale crack images of open datasets are not sufficient for detection of long and thin cracks of super large-scale image. The study proposes an automatic crack detection method for super large-scale images， based on feature pyramid network. Through four different feature maps in various sizes， the proposed network yields predictions， respectively， which means a highly precise crack segmentation. Experiments are carried on 120 steel box girder crack images in a resolution of 3 264 pixels×4 928 pixels. These images are used to train and test the network. The comparison study is conducted between the proposed method and the models trained with crack images resized into 1 600 pixels×2 400 pixels and 2 112 pixels×3 168 pixels with bilinear interpolation algorithm. The results show that our method can achieve the highest crack Intersection over Union （IoU） of 0.78， and the lowest Dice Loss of 0.12 in the comparison study. The predictions of crack images in testing indicate that resizing images sometimes result in the loss of crack information， and our method can maintain the detail of cracks and detect cracks of super large-scale images automatically and precisely.
转载注明来源:https://www.xzbu.com/8/view-15424492.htm