基于改进U-Net的卫星图像分割算法  

Improved U-Net segmentation algorithm of satellite image

作　　者：杨崎 张卓然 何嘉[1] YANG Qi;ZHANG Zhuo-ran;HE Jia(School of Computer Science,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学计算机学院,四川成都610225

出　　处：《计算机工程与设计》2023年第6期1714-1720,共7页Computer Engineering and Design

基　　金：四川省科技厅基金项目(SYZ202068)。

摘　　要：为解决传统模型与算法对遥感卫星图像小目标的分割精度低、泛化能力差等问题,提出一种基于改进U-Net的图像分割算法。将骨干网络改为ResNet18并加入优化后的空洞卷积池化金字塔与卷积注意力机制模块,充分提取小目标边缘特征。该算法在中国南部某地区的公开卫星图像数据集上的平均交并比与分割总精度分别达到了75.8%与95.6%,均超过U-Net、DeepLabV3+、SegNet、W-Net等主流语义分割网络。实验结果表明,该算法能有效改善网络的预测精度与小目标的分割结果。To solve the problems of low segmentation accuracy and poor generalization ability of small targets in remote sensing satellite images using traditional methods,an image segmentation algorithm based on the improved U-Net was proposed.The backbone network was changed to ResNet18 and the optimized atrous spatial pyramid pooling(ASPP)and convolutional block attention module(CBAM)were added to extract the edge features of small targets.The average intersection over union and overall accuracy of the algorithm on a public satellite image dataset of a certain region in southern China reach 75.8%and 95.6%,respectively,surpassing mainstream semantic segmentation networks such as U-Net,DeepLabV3+,SegNet,W-Net.Experimental results show that the algorithm effectively improves the prediction accuracy and the segmentation results of small targets.

关 键 词：图像分割 卫星图像 语义分割 小目标 空洞卷积 深度学习 注意力机制 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]