优化纹理匹配的路面裂缝检测模型  

Optimizing texture matching for pavement crack detection models

作　　者：沈德争 于滨 SHEN Dezheng;YU Bin(School of Civil Engineering,Hebei University of Engineering,Handan 056038,China)

机构地区：[1]河北工程大学土木工程学院,河北邯郸056038

出　　处：《重庆理工大学学报(自然科学)》2025年第1期93-101,共9页Journal of Chongqing University of Technology:Natural Science

基　　金：国家电网有限公司科技项目(J2023055)。

摘　　要：深度学习裂缝检测在特征提取过程中易发生信息丢失和位置偏移,面临复杂纹理路面背景和阴影等噪声干扰时出现裂缝的误检和漏检。提出一种基于参考的超分辨率思想的裂缝纹理匹配方法。优化后的上下文注意力模块(global context block,GC block)增强特征图片中像素间的空间依赖,随后从上层编码层参考图像中借用高分辨率空间纹理来补偿低分辨率特征图像中的信息损失,构建用于路面裂缝检测的纹理匹配网络(texture matching network,TMNet)模型。在公开数据集Crack500、DeepCrack和自建数据集上对预测模型进行验证。结果表明:所提出的TMNet在2个公开数据集和自建数据集上MIoU分别达到78.35%、89.86%和77.68%,相较于其他对比网络,能在细节纹理恢复上拥有更好的结果。In recent years,road maintenance has gained keen academic interest.The detection of pavement damage is a key part in road maintenance.Existing deep learning crack detection has problems such as information loss and positional offset in the process of feature extraction.Thus,misdetection and omission of cracks often occur when noise interference such as complex texture pavement background and shadows exists.We propose a crack texture matching method based on reference-based super-resolution idea.The optimized global context block(GC block)enhances the spatial dependence between pixels in the feature image and borrows high-resolution spatial texture from the reference image of the upper coding layer to compensate the information loss in the low-resolution feature image.Then,it builds the Texture Matching Network(TMNet)for the detection of cracks in pavement.The prediction model is validated on the public datasets(Crack500,DeepCrack)and our independently built dataset.Our results show the TMNet reaches 78.35%and 89.86%in MIoU on the two public datasets and 77.68%on our own dataset,demonstrating it performs better in detail texture recovery compared to other networks.

关 键 词：路面裂缝检测 语义分割 参考的超分辨率 注意力机制 纹理匹配 

分 类 号：U418.6[交通运输工程—道路与铁道工程]