基于FCC-Deeplabv3+的城市地下管道缺陷语义分割方法  

Sematic segmentation method for urban underground pipeline defects based on FCC-Deeplabv3+

作　　者：田淙文 李波[1] 蓝雯飞[1] 潘禹欣 姚为[1] TIAN Congwen;LI Bo;LAN Wenfei;PAN Yuxin;YAO Wei(College of Computer Science,South-Central Minzu University,Wuhan 430074,China)

机构地区：[1]中南民族大学计算机科学学院,武汉430074

出　　处：《中南民族大学学报（自然科学版）》2025年第1期107-117,共11页Journal of South-Central Minzu University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61976226)。

摘　　要：城市地下管道图像缺陷具有种类多、背景复杂、噪声多、缺陷尺度变化大等特点,导致目前城市地下管道缺陷分割算法精度不够高.本研究提出了一种基于Deeplabv3+的改进分割模型FCC-Deeplabv3+,并将该模型首次应用到城市地下管道缺陷分割.结合十字交叉注意力机制,使模型在预测时获取更丰富的上下文信息;提出了改进的解码器上采样策略,引入多尺度信息,减少中间层信息的丢失;使用基于增强的对比学习策略监督模型,提升了模型分割能力.此外,针对目前城市地下管道缺陷分割领域没有公开数据集的情况,基于Sewer-ML公开数据集,进行数据标注工作,构建了包含900张用于缺陷分割任务的数据集.通过实验验证了提出的缺陷分割模型的有效性及实时性,对比原始Deeplabv3+模型,mIoU提升了3.73%,mPA也提升了1.67%,并且相比其他基于深度学习的语义分割算法,也具有一定优势.The image defects of urban underground pipelines have the characteristics of multiple types,complex background,high noise,and large scale changes,which lead to insufficient accuracy of current urban underground pipeline defect segmentation algorithms.This research proposes an improved segmentation model FCC-Deeplabv3+based on Deeplabv3+and applies this model for the first time to defect segmentation of urban underground pipelines.Combined with the criss-cross attention mechanism,the method can obtain richer context when making predictions.An improved decoder upsampling strategy is proposed to introduce multi-scale information to reduce the loss of intermediate layer information.The method is supervised based on the contrastive learning strategy,which improves the method segmentation capability.In addition,regarding the current situation that there is no publicly available dataset for defect segmentation in urban underground pipelines,based on the Sewer-ML dataset,we performed data annotation work and constructed a dataset containing 900 images for the defect segmentation.The effectiveness and real time of the proposed defect segmentation method was verified through experiments.Compared with the original Deeplabv3+model,mIoU increased by 3.73%,and mPA also increased by 1.67%.It also has certain advantages compared with other semantic segmentation methods based on deep learning.

关 键 词：FCC-Deeplabv3+算法 缺陷分割 城市地下管道 十字交叉注意力 对比学习 深度监督 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]