基于改进Deeplabv3+的电力线分割方法研究
Research on power line segmentation method based on improved Deeplabv3+
唐心亮 1赵冰雪 1韩明 2宿景芳1
作者信息
- 1. 河北科技大学信息科学与工程学院 石家庄 050018
- 2. 石家庄学院未来信息技术学院 石家庄 050035
- 折叠
摘要
针对已有的分割算法存在的复杂场景干扰大、分割不准确的问题,提出一种用于电力线分割任务的改进Deeplabv3+模型.将原始主干网络替换为轻量级 Mobilenetv2网络,增加低水平特征,获得5路输入特征,充分提取特征信息;添加空洞空间金字塔池化(atrous spatial pyramid pooling,ASPP)的卷积分支数量,调整空洞率,提升图像的特征抓取能力,进一步在每个空洞卷积后加入1×1卷积操作,加快计算速度;提出一种基于坐标注意力机制的语义嵌入分支模块(coordinate attention semantic embedding branch,CASEB),融合第2、3路特征,增强目标特征的表示;引入卷积注意力机制模块(convolution block attention module,CBAM)抑制无用信息的传递,提高模型识别效率.实验结果表明,相对于原Deeplabv3+模型,改进模型在平均像素精度(mean pixel attention,MPA)和平均交并比(mean intersection over union,mIoU)上分别提升2.37%和3.42%,该方法可提供更加精确的电力线分割结果.
Abstract
In order to solve the problems of complex scene interference,inaccurate segmentation and slow prediction,an improved Deeplabv3+ model for power line segmentation is proposed.Replace the original backbone network with lightweight Mobilenetv2 network,add low-level features,obtain five-way input features,and fully extract feature information.The number of convolution in the atrous spatial pyramid pooling(ASPP)is increased,and the voidness rate was adjusted to improve the feature capturing ability of the image.Furthermore,1×1 convolution operation was added after each void convolution to speed up the calculation.A coordinate attention semantic embedding branch(CASEB)based on coordinate attention mechanism is proposed,which integrates the second and third features to enhance the representation of target features.CBAM attention module is introduced to inhibit the transmission of useless information and improve the efficiency of model recognition.Compared with the original Deeplabv3+ model,the mean pixel attention(MPA)and mean intersection over union(mIoU)of the improved model are improved by 2.37%and 3.42%respectively.This method can provide more accurate results of power line segmentation.
关键词
电力线分割/深度学习/改进Deeplabv3+模型/Mobilenetv2/注意力模块Key words
power line segmentation/deep learning/improved Deeplabv3+ model/Mobilenetv2/attention module引用本文复制引用
基金项目
石家庄市科技计划项目(221130321A)
出版年
2024
NETL
NSTL
维普
万方数据