基于多尺度特征图的变电站设备缺陷递进检测

张可1; 2; 黄文礼1; 2; 侯仕杰1; 2; 代北敏3; 陈思宝3

基于多尺度特征图的变电站设备缺陷递进检测

引用本文：张可1,2,黄文礼1,2，侯仕杰1,2，代北敏3，陈思宝3.基于多尺度特征图的变电站设备缺陷递进检测[J].计算技术与自动化,2025,(2):134-141

摘要点击次数: 1013

全文下载次数: 747

作者	单位
张可1,2,黄文礼1,2，侯仕杰1,2，代北敏3，陈思宝3	(1. 安徽南瑞继远电网技术有限公司,安徽合肥 230088；2. 国网电力科学研究院有限公司（南瑞集团有限公司）,江苏南京 211000；3. 安徽大学计算机科学与技术学院,安徽合肥 230601)

中文摘要:针对现有变电站设备缺陷检测中目标尺度差异大、背景复杂等问题，提出了一种基于多尺度特征图的变电站设备缺陷递进检测方法。首先通过深度残差网络提取变电站设备缺陷图像的原始特征图，利用多尺度特征聚合方法构建多尺度特征图；其次对缺陷目标进行两阶段的递进检测，第一阶段对缺陷目标进行前景和背景的二分类与坐标回归，第二阶段对前景目标的具体缺陷类别进行判断和进一步坐标回归。本文方法同时将第一阶段的类别结果作为语义信息对特征图添加空间注意力机制，能够对不同尺度的特征图的特征值进行自适应加权，增强前景区域的特征响应，同时抑制背景区域，从而增强了缺陷检测方法的鲁棒性和泛化能力。在真实变电站设备缺陷图像数据集上的实验结果表明所提方法具有良好的检测性能。

中文关键词:缺陷检测目标检测递进检测变电站设备多尺度特征图

Substation Equipment Defect Progressive Detection Based on Multi scale Feature Map

Abstract:Aiming at the problem of poor robustness of existing detection methods when object scale is variant and background is complex, this paper proposes a progressive detection method for substation equipment defects based on multi scale feature maps. Firstly, original feature maps of substation equipment defect images are extracted by deep residual network, and multi scale feature maps of substation equipment images are constructed by multi scale feature aggregation method. Secondly, the detection process goes through two stages. The first stage is the classification of foreground and background of defect objects and coordinate regression. The second stage determines the specific category of foreground objects and further performs coordinate regression. The first stage category results are used as semantic information to add a spatial attention mechanism on the feature map, which can adaptively weight the feature values of feature maps of different scales to enhance the feature response of the foreground area, while suppressing the background area, thereby enhancing the robustness and generalization ability of the proposed method. Experimental results on real substation equipment defect image data set show that the proposed method has better detection performance.

keywords:defect detection object detection progressive detection substation equipment multi scale feature map

查看全文 查看/发表评论 下载pdf阅读器