基于深度学习与局部感知注意力的避雷器套管红外图像分割方法  

Infrared Image Segmentation Method of Lightning Arrester Casing Based on Deep Learning and Local Perception Attention Mechanism

作　　者：王生祺 陈海波 叶金翔 WANG Shengqi;CHEN Haibo;YE Jinxiang(State Grid Zhejiang Electric Power Company Limited Ultra High Voltage Branch,Hangzhou 310000,China)

机构地区：[1]国网浙江省电力有限公司超高压分公司,浙江杭州310000

出　　处：《湖南电力》2025年第2期122-128,共7页Hunan Electric Power

摘　　要：针对现有避雷器套管实例分割方法精度低、容易误检的问题,提出一种基于深度学习的避雷器套管分割检测方法,通过经典神经网络模型UNet实现端到端的分割推理,同时,设计局部感知注意力模块高效聚合避雷器套管红外图像的通道信息与空间信息,避免红外背景噪声影响,实现边缘清晰的分割效果。通过定性和定量的实验比较验证模型UNet在避雷器套管红外图像分割检测的性能。结果表明,该模型对避雷器套管具有较好的检测分割效果,对后续故障检测、降低劳动力成本、保证变电站设备安全运行具有重要意义。Addressing the issues of low accuracy and susceptibility to false detection in the existing instance segmentation methods for lightning arrester casing,a deep learning-based lightning arrester casing segmentation and detection method is proposed to achieve end-to-end segmentation inference through the classic neural network model UNet.Additionally,the designed local perception attention module efficiently aggregates the channel and spatial information of the lightning arrester casing infrared image,overcoming the impact of infrared background noise and achieving clear edge segmentation effect.The performance of the model Unet in infrared image segmentation and detection of lightning arrester casings is validated through qualitative and quantitative experimental comparisons.The results show that this model offers superior detection and segmentation performance for lightning arrester casings,which is significant for subsequent fault detection,reducing labor costs,and ensuring the safe operation of substation equipment.

关 键 词：避雷器套管 红外图像 分割检测 计算机视觉 深度学习 注意力机制 

分 类 号：TM862.1[电气工程—高电压与绝缘技术] TP391.4[自动化与计算机技术—计算机应用技术]