全文下载:
202506018.pdf
文章编号: 1672-6987(2025)06-0142-09 DOI: 10.16351/j.1672-6987.2025.06.018
张孔增1, 李晓晖2, 侯竣凯2, 曹怡亮2, 边太成2(1.菏泽市定陶区房产服务中心, 山东 菏泽 274100;2.青岛科技大学 信息科学技术学院, 山东 青岛 266061)
摘要: 由于遥感图像的分布不同,预训练的分割模型在新数据集上难以保持原有性能,尤其是在无标签数据上。此外,大型数据集的像素级标签所需成本较大。针对此问题,本文提出了一种基于双向学习的遥感图像域自适应语义分割方法(ASBLNet)。该方法利用双向学习将图像翻译模型与分割模型相结合,进行端到端的联合训练,利用分割模型的结果来改进翻译模型,最终利用改进后的翻译模型来优化分割模型,实现相互促进的目的。相较于传统方法,该方法实现了翻译与分割模型相互促进的目的。实验结果表明,本文所提出的方法在两个公共遥感数据集PotsdamIRRG和Vaihingen上所取得的分割结果优于现有的方法,证明了该方法的有效性。
关键词: 域自适应; 双向学习; 遥感图像; 语义分割
引用格式: 张孔增, 李晓晖, 侯竣凯, 等. 基于双向学习的遥感图像域自适应语义分割方法[J]. 青岛科技大学学报(自然科学版), 2025, 46(6): 142-150.
ZHANG Kongzeng, LI Xiaohui, HOU Junkai, et al. Adaptive semantic segmentation method based on bidirectional learning in remote sensing image domain[J]. Journal of Qingdao University of Science and Technology(Natural Science Edition), 2025, 46(6): -.
Adaptive Semantic Segmentation Method Based on Bidirectional Learning in Remote Sensing Image Domain
ZHANG Kongzeng1, LI Xiaohui2, HOU Junkai2, CAO Yiliang2, BIAN Taicheng2
(1.Real Estate Service Center of Dingtao District, Heze 274100, China;
2.College of Information Science and Technology, Qingdao University of Science and Technology, Qingdao 266061, China)
Abstract: Due to the differing distributions of remote sensing images, pre-trained segmentation models struggle to maintain their original performance on new datasets, particularly on unlabeled data. Additionally, the cost of pixel-level labeling for large datasets is substantial. To address this issue, this paper proposes a domain a adaptive semantic segmentation method based on bidirectional learning in remote sensing image domain(ASBLNet). This method integrates an image translation model with a segmentation model through bidirectional learning, enabling end-to-end joint training. The results from the segmentation model are used to refine the translation model, which in turn is utilized to optimize the segmentation model, achieving mutual enhancement. Compared to traditional methods, this approach facilitates mutual improvement between the translation and segmentation models. Experimental results demonstrate that the proposed method achieves superior segmentation outcomes on two public remote sensing datasets, PotsdamIRRG and Vaihingen, outperforming existing methods. This validates the effectiveness of the proposed approach.
Key words: domain adaptation; bidirectional learning; remote sensing images; semantic segmentation
收稿日期: 2024-12-18
基金项目: 山东省产教融合研究生联合培养示范基地项目(2020-19).
作者简介: 张孔增(1975—),男,工程师.