基于改进DeeplabV3+的遥感图像道路分割模型
Road segmentation model of remote sensing image based on improved DeeplabV3+
张银胜 1单梦姣 2钟思远 3陈戈 2童俊毅 2单慧琳1
作者信息
- 1. 南京信息工程大学电子与信息工程学院 南京 210044;无锡学院电子信息工程学院 无锡 214105
- 2. 南京信息工程大学电子与信息工程学院 南京 210044
- 3. 无锡学院电子信息工程学院 无锡 214105
- 折叠
摘要
针对遥感图像道路分割边界模糊和遮挡难以区分的问题,提出了基于改进DeeplabV3+的遥感图像道路分割模型.该模型在主干网络中引入 MobileNetV3和高效通道注意力机制(ECA),减少了参数量并关注连续的道路特征信息.在解码过程中采用多级上采样,增强了编码器和解码器之间的紧密连接,全面保留了细节信息.同时,在ASPP模块中采用深度可分离膨胀卷积DS-ASPP,显著减少了参数量.实验结果表明,该模型在 Massachusetts Roads数据集上的交并比达到了83.71%,准确率达到了93.71%,分割精度最优,模型参数量为55.57×106,能够有效地避免边界模糊和遮挡导致的错漏检问题,在遥感道路分割中提高了精度和速度.
Abstract
In response to the problem of fuzzy boundaries and difficulty in distinguishing occlusions in road segmentation of remote sensing images,this paper proposes a remote sensing image road segmentation model based on an improved DeeplabV3+.The model introduces MobileNetV3 and ECA attention mechanism in the backbone network to reduce parameter volume and focus on continuous road feature information.In the decoding process,multi-level upsampling is adopted to enhance the tight connection between the encoder and decoder,fully preserving detailed information.Meanwhile,deep separable dilated convolution(DS-ASPP)is used in the ASPP module to significantly reduce the number of parameters.The experimental results demonstrate that the model achieves an intersection over union(IoU)of 83.71%and an accuracy of 93.71%on the Massachusetts Roads dataset.With a parameter count of 55.57×106,the model exhibits superior segmentation accuracy and effectively avoids errors and omissions caused by boundary blurring and occlusion.It enhances both precision and speed in remote sensing road segmentation.
关键词
遥感图像/道路分割/DeeplabV3+模型/MobileNetV3模型/多级上采样Key words
remote sensing images/road segmentation/DeeplabV3+model/mobileNetV3 model/multi-level upsam-pling引用本文复制引用
基金项目
国家自然科学基金(62071240)
江苏省产教融合型一流课程(2022-133)
无锡学院2023年教改研究课题(XYJG2023010)
无锡学院2023年教改研究课题(XYJG2023011)
出版年
2024
NETL
NSTL
万方数据