基于感兴趣区域约束的漏缆卡扣重构缺陷检测  

Detecting reconstruction defects in leaky cable snaps based onregion of interest constraints

作　　者：任帅 李毅 肖武龙 王逸涵 李柏林[1] Ren Shuai;Li Yi;Xiao Wulong;Wang Yihan;Li Bailin(School of Mechanical Engineering,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学机械工程学院,成都610031

出　　处：《电子测量技术》2024年第20期167-176,共10页Electronic Measurement Technology

基　　金：四川省科技研发计划重点项目(2021YFN0020)资助。

摘　　要：针对自监督漏缆卡扣缺陷算法中因背景重构误差造成的误检测,本文提出一种基于感兴趣区域约束的双阶段漏落卡扣缺陷重构方法。针对目标检测算法定位到的分类卡扣区域图像,首先嵌入分割网络用于分离卡扣及漏缆区域。随后,其对应掩码被作为线索引导堆叠对抗生成网络对卡扣区域进行重构,保证背景一致性的同时,精细化重构卡扣缺陷区域。此外,通过嵌入深度残差块和改进损失函数迫使生成网络更专注于感兴趣区域的重构。最后,训练完成的网络被用于卡扣图像的重构,并根据重构前后图像的相似度分数判定是否存在缺陷。在漏缆卡扣数据集上,定量结果表明本文算法对多类型卡扣缺陷识别的准确率92.3%,召回率为93.4%,高于其他自监督卡扣重构方法。可视化结果表明所提卡扣重构方法背景重构误差较少。Aiming to address the false detections caused by background reconstruction errors in the self-supervised leaky cable snap defect detection algorithm,this paper proposes a two-stage leaky drop snap defect reconstruction method based on region of interest constraints.For the classified snap region images localized by the target detection algorithm,a segmentation network is first employed to differentiate between snap and cable leakage regions.Subsequently,the corresponding masks are utilized to guide a stacked adversarial generative network to reconstruct the snap regions,ensuring high-quality reconstruction of the defect areas while maintaining background consistency.Additionally,the generative network is optimized to place greater emphasis on reconstructing the regions of interest by integrating deep residual blocks and refining the loss function.Ultimately,the trained network is deployed for the reconstruction of snap images,determining the presence of defects based on the similarity scores of the images before and after reconstruction.Quantitative results on the leaky cable snap dataset demonstrate that the proposed algorithm achieves a defect recognition accuracy of 92.3%and a recall rate of 93.4%,surpassing the performance of other self-supervised snap reconstruction methods.Visualization results further indicate a reduction in background reconstruction errors in the proposed method.

关 键 词：漏缆卡扣 缺陷检测 感兴趣区域引导 堆叠对抗生成网络 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] U285.6[自动化与计算机技术—计算机科学与技术] TN913[交通运输工程—交通信息工程及控制]