基于深度可分离空洞卷积金字塔的变压器渗漏油检测

Detection of transformer oil leakage based ondeep separable atrous convolution pyramid

为了降低影响并提高对变压器渗漏油巡检图像的检测效率,提出一种基于深度可分离空洞卷积金字塔的变压器渗漏油检测模型。首先,将空洞金字塔中普通卷积块修改为深度可分离卷积块,以此扩大金字塔感受野,使特征提取网络提取到的特征图语义信息更加丰富;然后,改进了特征提取阶段低阶语义特征与高阶语义特征融合过程,进一步增强特征提取网络产生特征图的语义信息;最后,为了避免经过多次卷积、池化操作后特征图语义信息的损失,在融合过程中引入空间注意力机制和通道注意力机制,进一步增强特征图中的语义信息。与UNet(convolutional networks for biomedical image segmentation)、PSPNet(pyramid scene parseing network)、DeepLabv3+(encoder-decoder with atrous separable convolution for semantic image segmentation)和MCNN(multi-class convolutional neural network)等算法进行对比实验发现,本文所提出网络检测模型效果好,查准率达到了76.85%,平均交并比达到了64.63%,召回率达到了73.56%,检测速率达到了30 f/s。为了验证本文提出方法的有效性,设计了消融实验,与基础网络模型相比,查准率提高了9.33%,平均交并比提高了7.15%,召回率提高了5.66%。

To improve the detection efficiency of the transformer oil leakage patrol inspection image,a deep separable atrous convolution pyramid-based transformer oil leakage detection model is suggested.First,the ordinary convolution block in the atrous pyramid is modified into a deep separable convolution block for expansion of the pyramid receptive field and further enrichment of the semantic information of the feature graph extracted by the feature extraction network.Afterward,the fusion of low-order and high-order semantic features in the feature extraction stage is improved for fur-ther enhancement of the semantic information of the feature graph generated by the feature extraction network.Finally,to avoid semantic information loss in the feature graph after several convolution and pooling operations,spatial atten-tion and channel attention mechanisms are introduced into the fusion process to further enhance the semantic informa-tion in the feature graph.It is found by comparing with algorithms such as traditional UNet(Convolutional Networks for Biomedical Image Segmentation),PSPNet(Pyramid Scene Parsing Network),DeepLabv3+(Encoder-Decoder with At-rous Separable Convolution for Semantic Image Segmentation),and MCNN(Multiclass Convolutional Neural Network)via tests that the proposed network detection model is effective,with 76.85%precision,64.63%average cross-merger ratio,73.56%recall rate,and 30 frames per second.To confirm the effectiveness of the proposed method,an ablation experiment is designed.Compared with the basic network model,the precision,average intersection ratio,and recall rate are increased by 9.33%,7.15%,and 5.66%,respectively.