基于深度学习的裂纹实例分割方法  

Crack instance segmentation method based on deep learning

作　　者：李棋 路胜男 马千里[1] LI Qi;LU Shengnan;MA Qianli(School of Communications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学通信与信息工程学院,南京210003

出　　处：《智能计算机与应用》2024年第12期18-25,共8页Intelligent Computer and Applications

摘　　要：本文针对传统裂纹检测方法的检测精度较低的问题,提出一种先验知识改进和深度学习相融合的裂纹实例分割方法。使用单阶段网络YOLOv5进行实验验证,先验知识改进包括使用裂纹图像分类模型过滤mosaic增强产生的错误样本并对相连接裂纹的标签信息进行融合,实现训练集的纠正与增强;使用基于交并比的K-means++算法生成先验框,加速训练回归收敛。最终实验验证,先验知识改进不增加网络参数,实现更高的检测精度,同时还保持单阶段网络较高的检测速度,验证了对于先验知识的改进可以有效地提高裂纹实例分割的检测精度,增强网络模型的泛化性。Aiming at the problem of low detection accuracy of traditional crack detection methods,this paper proposes a crack instance segmentation method which integrates prior knowledge improvement and deep learning.The single-stage network YOLOv5 is used for experimental verification.The prior knowledge improvement includes the use of crack image classification model to filter the error samples generated by mosaic enhancement and the fusion of the label information associated with the connected cracks to realize the correction and enhancement of the training set.The K means++algorithm based on intersection ratio is used to generate prior frames and accelerate the training regression convergence.The final experiment verifies that the improvement of prior knowledge does not increase network parameters,achieving higher detection accuracy,while maintaining a high detection speed of single-stage network,verifying that the improvement of prior knowledge can effectively improve the detection accuracy of crack instance segmentation and enhance the generalization of network model.

关 键 词：裂纹检测 实例分割 mosaic增强 先验框 K-means++ 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]