改进SSD的变电站电力设备识别方法

Improved SSD Detection Method for Substation Power Equipment

基于机器视觉的变电站设备识别是电力智能巡检的关键环节,在巡检图像上不同设备尺寸差异显著,极大影响了现有算法的识别性能。针对尺寸差异显著的识别对象,提出一种改进的SSD算法用于变电站设备识别任务。该方法主要改进SSD算法的两个重要部分:其一,保证多尺度特征提取,加强大尺度特征图的有效性。其二,优化默认框的生成策略,提出符合尺寸差异显著对象的多种默认框。此外,为增强网络的鲁棒性,采用迁移学习技术和数据增强技术训练电力设备识别网络模型。最后,构建变电站电力设备数据集,并进行实验测试;实验结果表明,论文方法相较于传统SSD算法,召回率提升了42.24%,解决了尺寸差异显著对象的识别问题。

The detection of substation equipment based on machine vision is a key link in the intelligent inspection of electric power.The size of different equipment in the inspection images is significantly different,which greatly affects the detection perfor-mance of the existing algorithms.Aiming at the detection objects with significant size differences,an improved SSD algorithm is pro-posed for substation equipment detection tasks.This method mainly improves two important parts of the SSD algorithm.First,it en-sures the extraction of multi-scale features and strengthens the effectiveness of large-scale feature maps.Second,it optimizes the generation strategy of the default frame,and proposes a variety of default frames that meet the objects with significant differences in size.In addition,in order to enhance the robustness of the network,transfer learning technology and data enhancement technology are used to train the power equipment detection network model.Finally,experiments are carried out on the data set of power equip-ment in substations,and the method obtains the best accuracy and recall rate,which shows that it can better solve the problem of de-tecting objects with significant size differences.