摘要
提高电气设备紫外图像分割精确度对设备放电程度的准确评估具有重要意义。由于存在噪声干扰与紫外光斑形状、大小不规则等问题,目标分割区域存在过分割和欠分割现象,因此提出一种基于多模块的VSA-UNet(VGG16Net, Improved SENet, and ASPP based U-Net)分割网络。为强化网络特征提取能力,减少过分割现象,使用VGG16Net的卷积层代替U-Net网络的编码部分;将编码部分末端卷积层替换成空洞空间金字塔池化(Atrous Spatial Pyramid Pooling, ASPP)模块,获取紫外图像的多尺度信息,解决大区域的欠分割问题;在跳跃连接部分加入改进SENet模块,加强有用信息的提取,补充细节损失,提升整体网络性能。基于自建紫外图像数据集的实验表明,改进网络在分割紫外图像时平均交并比(Mean Intersection over Union, MIoU)达到81.78%,平均精确率为95.97%。与U-Net网络相比,提出的VSA-UNet模型明显提升了紫外图像分割的准确性。
Improving the accuracy of ultraviolet image segmentation of electrical equipment is of great significance to the accurate evaluation of the discharge degree of the equipment. Due to the problems of noise interference and irregular shape and size of ultraviolet light spots, the target segmentation area has over-segmentation and under-segmentation. Therefore, a multi-module-based VSA-UNet(VGG16Net, Improved SENet, and ASPP based U-Net) is proposed to split the network. Firstly, in order to strengthen the network feature extraction capability and reduce the phenomenon of over-segmentation, the convolutional layer of VGG16Net is used to replace the encoding part of the U-Net network. Secondly, the end convolution layer of the encoding part is replaced with an Atrous Spatial Pyramid Pooling(ASPP) module to obtain multi-scale information of ultraviolet images and solve the problem of under-segmentation in large areas. Finally, an improved SENet module is added to the skip connection part to enhance the extraction of useful information, supplement the loss of details, and improve the overall network performance. Experiments based on the self-built ultraviolet image dataset show that the Mean Intersection over Union(MIoU) of the improved network in segmenting ultraviolet images reaches 81.78%, and the average accuracy is 95.97%. Compared with the U-Net network, the proposed VSA-UNet model significantly improves the accuracy of UV image segmentation.
作者
陈思林
秦伦明
王悉
杨苏航
左安全
CHEN Silin;QIN Lunming;WANG Xi;YANG Suhang;ZUO Anquan(College of Electronics and Information Engineering,Shanghai University of Electric Power,Shanghai 201306,China;School of Electronic and Information Engineering,Beijing Jiaotong University,Beijing 100044,China)
出处
《无线电工程》
北大核心
2023年第1期230-238,共9页
Radio Engineering
基金
国家自然科学基金(62073024)。