嵌入视觉关系掩码的多类别金具检测方法  

A multi-category fitting detection method with embedded visual relation masks

作　　者：王巍 杨耀权[1] 王乾铭 翟永杰[1] 赵振兵[2] WANG Wei;YANG Yaoquan;WANG Qianming;ZHAI Yongjie;ZHAO Zhenbing(Department of Automation,North China Electric Power University,Baoding 071003,China;Department of Electronic and Communication Engineering,North China Electric Power University,Baoding 071003,China)

机构地区：[1]华北电力大学自动化系,河北保定071003 [2]华北电力大学电子与通信工程系,河北保定071003

出　　处：《智能系统学报》2023年第3期440-449,共10页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(U21A20486,61871182);河北省自然科学基金项目(F2021502008).

摘　　要：在对输电线路金具进行检测的过程中,由于受到复杂背景的影响,一些互相遮挡或者特征不明显的金具会隐匿在复杂环境难以精确检测。针对这一问题,提出了基于视觉关系掩码的多类别金具检测模型,通过挖掘和提取输电线路金具之间包含空间信息的视觉关系先验知识,构建视觉关系掩码和视觉关系检测网络,并将先验知识作为辅助信息融入视觉关系模块中,最终实现多类别金具的精确定位与识别。对具有14类金具的数据集进行多种定性和定量实验,结果表明,改进后的模型平均检测精度能提高到76.25%,检测效果也优于其他先进目标检测模型。In the process of detecting power transmission line fittings,due to the influence of complex backgrounds,some fittings with mutual obscuration or inconspicuous features will be hidden in complex environments and it is hard to detect them accurately.In response to this problem,this paper proposes a multi-category hardware detection model based on visual relationship masks.By mining and extracting the prior knowledge of visual relationships between power transmission line fittings that contain spatial information,the visual relationship mask(VRM),and visual relationship detection network(VRDN)are constructed,and the prior knowledge is integrated into the visual relationship module as auxiliary information,realizing the precise positioning and recognition of multi-category fittings.A variety of qualitative and quantitative experiments have been performed on a data set with 14 types of hardware.The results show that the average detection accuracy of the improved model can be increased to 76.25%,and the detection effect is better than other advanced target detection models.

关 键 词：目标检测 输电线路 金具 深度学习 视觉关系 先验知识 空间信息 辅助信息 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]