无线电工程2025，Vol.55Issue(12)：2362-2372,11.DOI:10.3969/j.issn.1003-3106.2025.12.005
基于DeepLabv3+改进的轻量化多尺度特征融合遥感图像分割网络
Lightweight Multi-scale Feature Fusion Remote Sensing Image Segmentation Network Based on DeepLabv3+
摘要
Abstract
To address the issues of large amount of model parameters,complex model structure,and high training resource consumption in traditional remote sensing image semantic segmentation,a lightweight multi-scale feature fusion remote sensing image semantic segmentation network—Lightweight Multi-scale Feature Fusion Network(LMFFNet)based on the improvement of DeepLabv3+is proposed,aiming to reduce the model complexity and number of parameters while maintaining the accuracy of semantic segmentation results.The network,which is based on DeepLabv3+,adopts MobileNetV2 at the encoder end as the backbone network to reduce the number of model parameters,and the Efficient Atrous Spatial Pyramid Pooling(EASPP)module is used to replace the Atrous Spatial Pyramid Pooling(ASPP)module,maintaining the multi-scale receptive field while reducing the model complexity.At the decoder end,an Adaptive Cross-level Feature Fusion Module(ACFM)based on the Squeeze-and-Excitation Network(SENet)attention mechanism is proposed to achieve adaptive dynamic adjustment of feature fusion weights and capture features from different levels more effectively.Transposed convolution is used in the upsampling process to optimize image reconstruction quality.An Efficient Channel Attention Network(ECA-Net)lightweight attention mechanism module is used for more efficient channel feature interaction and fusion.Experiments conducted on the Potsdam and Vaihingen datasets indicate that the mean Intersection of Union(mIoU)of LMFFNet reaches 70.26％and 69.77％respectively,with a model parameter size of only 16.45 MB.The experimental results show that compared with mainstream network models such as UNet,PSPNet,and DeepLabv3+,LMFFNet achieves a lightweight model while ensuring the accuracy of remote sensing image semantic segmentation.关键词
遥感图像/语义分割/轻量化/特征融合/注意力机制Key words
remote sensing images/semantic segmentation/lightweight/feature fusion/attention mechanism分类
信息技术与安全科学引用本文复制引用
卢洪轩,徐爱茹,姚兴兴,蒋林烽,张耀严,余子怡..基于DeepLabv3+改进的轻量化多尺度特征融合遥感图像分割网络[J].无线电工程,2025,55(12):2362-2372,11.基金项目
武汉工程大学第十八期大学生校长基金(XZJJ2024043) （XZJJ2024043）
武汉工程大学第十六届研究生教育创新基金(CX2024088)18th President Fund of Wuhan Institute of Technology(XZJJ2024043) （CX2024088）
16th Postgraduate Education Innovation Fund of Wuhan Insti-tute of Technology(CX2024088) （CX2024088）
