基于重参数化和联合分支的城市地下管道缺陷检测

Defect detection for urban underground pipelines based on reparameterization and joint branch

城市地下管道是城市重要的基础设施之一,及时排查管道缺陷对城市的发展起着较为重要的作用,针对目前的管道缺陷检测模型参数量大、实时性较差等问题,提出一种改进的FCOS城市地下管道缺陷检测方法.首先,引入轻量的MobileOne网络,通过结构重参数化将多分支网络转换为单分支网络,减小模型规模;然后引入分类和IoU的联合分支使模型的训练和推理过程保持一致,并利用平衡因子优化QFL损失函数,提升模型分类预测效果.实验结果表明:改进后的FCOS模型相比于基线模型的平均精度提升1.83%,检测速度FPS达到48.6,模型参数量下降17.85 M,有效提升了城市地下管道缺陷检测性能,并且相比于其他优秀的目标检测算法,也具有一定的优势.

Urban underground pipeline is one of the important infrastructures in the city,and the timely detection of pipeline defects plays a more important role in the development of the city.Aiming at the current pipeline defect detection model with large parameter number and poor real-time performance,an improved FCOS method of the defect detection is proposed for urban underground pipelines.Firstly,the lightweight MobileOne network is introduced,and the model size is reduced by converting the multi-branch network into a single-branch network through structural reparameterization.Then,the joint branch of classification and IoU is introduced to make that the model's training is consistent with inference process,and the balancing factor is utilized to optimize the QFL,which improves the classification prediction effect of the model.The experimental results show that the improved FCOS model improves the average accuracy by 1.83%compared with the baseline model,the detection speed FPS reaches 48.6,and the number of model parameters decreases by 17.85 M,which effectively improves the performance of defects detection for urban underground pipelines,and it also has certain advantages compared with other excellent target detection algorithms.