VCDG-UNet模型在遥感图像分割中的应用  

Application of VCDG-UNet Model in Remote Sensing Image Segmentation

作　　者：郑海洋 于淼[1] 于晓鹏[1] ZHENG Haiyang;YU Miao;YU Xiaopeng(College of Mathematics and Computer,Jilin Normal University,Siping 136099,China)

机构地区：[1]吉林师范大学数学与计算机学院,吉林四平136099

出　　处：《无线电工程》2025年第1期94-104,共11页Radio Engineering

基　　金：吉林省科技发展计划项目(YDZJ202301ZYTS285)。

摘　　要：针对遥感图像建筑物的轮廓分割不完整、边界分割模糊和阴影干扰等导致的错误分割问题,提出一种基于VGG16的卷积块注意力深度可分离卷积U-Net网络(VGG16 Convolutional Block Attention and Deep Separable Convolution U-Net,VCDG-UNet)。为对建筑物特征进行提取,编码器部分模型以具有强大特征提取能力的VGG16作为骨干网络;解码器部分用深度可分离卷积代替普通卷积来减少参数量并融合不同尺度的特征;引入卷积块注意力模块(Convolutional Block Attention Module,CBAM)加入跳跃连接中,使其更有效地从不同尺度的图像中提取上下文信息并提高其对重要区域的关注度;为解决网络训练过程中的梯度消失问题,使用了高斯误差线性单元(Gaussian Error Linear Unit,GELU)。实验结果显示,改进后的网络在WHU和INRIA数据集上的平均交并比(mean Intersection over Union,mIoU)和F1-score分别达到了94.20%、96.83%和89.69%、94.51%,相较于基础模型高出了1.59%、0.76%和2.8%、1.59%。To solve the problem of incorrect segmentation caused by incomplete contour segmentation,blurred boundary segmentation,and shadow interference in remote sensing building image,an improved end-to-end convolutional neural network VGG16 Convolutional Block Attention and Deep Separable Convolution U-Net(VCDG-UNet)based on the U-Net encoder-decoder structure is proposed.In the encoder part,the model uses the VGG16 backbone network with strong feature extraction capability to extract building features.In the decoder part,replace the regular convolution with depthwise separable convolution to reduce the number of parameters and fuse features of different scales.The Convolutional Block Attention Module(CBAM)is introduced into the skip connections to more effectively extract context information from multi-scale images and increase the attention on important regions.The Gaussian Error Linear Unit(GELU)activation method is used to replace the RELU activation method to solve the problem of vanishing gradients during the network training.According to the experimental results,the improved network achieves mIoU and F1-score of 94.20%,96.83%and 89.69%,94.51%on the WHU and INRIA datasets respectively,which are 1.59%,0.76%and 2.8%,1.59%higher than the baseline model.

关 键 词：遥感图像分割 深度学习 U-Net 卷积块注意力模块 高斯误差线性单元 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]