基于VGG16-Unet算法的扭角式压板分割识别方法

Segmentation and Identification Method of Twisted Platen Based on VGG16-Unet Algorithm

针对当前变电站人工巡检压板状态工作效率低,且现有扭角式压板图像识别效果不佳的问题,提出了一种基于VGG16 Unet语义分割模型的压板状态识别方法。首先设计了VGG16-Unet的网络模型结构,该模型结构包含主干特征提取网络部分、加强特征提取网络部分和预测网络部分,在网络的下采样和上采样过程中捕捉图像深层次语义特征和浅层次细节特征;其次定义了网络模型的Dice损失函数并分析4种性能评估函数检测压板识别效果;最后在1000张压板数据集试验中,该深度学习方法精确率高达98.6%,召回率92.2%,综合指标95.3%,平均交并比92.4%,与现有主流压板状态识别方法进行对比分析,结果显示本文方法具有更好的识别性能。

In view of the low efficiency of manual inspection of the status of the transformer substation,and the poor effect of the image recognition of the existing torsional Angle clamp,a new method of the status recognition of the clamp based on the semantic segmentation model of VGG16-Unet was proposed.Firstly,the network model structure of VGG16-Unet is designed,which includes the trunk feature extraction network part,the enhanced feature extraction network part and the prediction network part,and captures the deep semantic features and shallow detail features of the image during the down-sampling and up-sampling of the network.Secondly,the Dice loss function of network model is defined and four performance evaluation functions are analyzed to detect the recognition effect of platen.In the final experiment on a dataset of 1000 pressure plates,the deep learning method achieved an accuracy of 98.6%,a recall rate of 92.2%,a comprehensive index of 95.3%,and an average intersection to union ratio of 92.4%.Compared with the existing mainstream pressure plate state recognition methods,the results show that the proposed method has better recognition performance.