小样本海水养殖SAR影像的两阶段生成对抗网络语义分割  

Two stage generative adversarial network semantic segmentation for few-shot mariculture SAR images

作　　者：邢军[1] 于航 王心哲 范剑超 XING Jun;YU Hang;WANG Xinzhe;FAN Jianchao(School of Information Science and Engineering,Dalian Polytechnic University,Dalian 116034,China;School of Control Science and Engineering,Dalian University of Technology,Dalian 116023,China)

机构地区：[1]大连工业大学信息科学与工程学院,辽宁大连116034 [2]大连理工大学控制科学与工程学院,辽宁大连116023

出　　处：《大连工业大学学报》2025年第1期64-72,共9页Journal of Dalian Polytechnic University

基　　金：国家自然科学基金项目(42076184)。

摘　　要：针对海水养殖遥感影像的标签标注工作人工成本高,传统语义分割技术难以学习养殖提取目标上下文信息的问题,设计了基于生成对抗网络的两阶段网络框架。第一阶段网络负责生成任务,使用SAR图像养殖区清晰度区分策略选取小样本数据,生成语义信息相似的无配对信息伪数据,解决循环一致生成对抗网络模型使用无配对信息海水养殖数据集时产生过拟合的问题;第二阶段网络负责语义分割任务,引入循环一致损失,使用第一阶段网络得到的伪数据和选取的小样本数据完成训练;最后设置阈值函数降低预测的像素值误差,提高语义分割精度。实验在GF-3数据集上分别与6种对比方法进行比较,总体精度为83.54%,平均交并比为0.7032,优于其他对比模型。The labeling of mariculture remote sensing images is time-consuming and labor-intensive,and traditional semantic segmentation technology is difficult to learn and extract target context information.In order to solve the above problems,the two-stage network framework based on generative adversarial networks was designed.The first stage network was responsible for generating pseudo-data,using the synthetic aperture radar(SAR)image definition differentiation strategy to select small sample data,and generating pseudo-data without paired information with similar semantic information,to solve the problem of overfitting when the cycle-consistent adversarial network(CycleGAN)model used the mariculture remote sensing images and labels without paired information.The second stage network was responsible for semantic segmentation,introduced cyclic consistent loss,and used the pseudo-data obtained from the first stage network and selected small sample data to complete training,reducing the cost of manual pixel-level labeling.Finally,the threshold function was set to reduce the predicted pixel value error and improve the semantic segmentation accuracy.The experiment was compared with six comparison methods on the GF-3 dataset,the overall accuracy was 83.54%and the mean intersection over union ratio was 0.7032,which was better than other comparison models.

关 键 词：海水养殖 合成孔径雷达 语义分割 小样本学习 生成对抗网络 循环一致性 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]