基于强弱一致性的无监督领域自适应语义分割方法  

Unsupervised domain adaptive semantic segmentation method based on strong-weak consistency

作　　者：田晓敏 郭仁春 赵怀慈[2] 刘鹏飞 房建 TIAN Xiaomin;GUO Renchun;ZHAO Huaici;LIU Pengfei;FANG Jian(School of Information Engineering,Shenyang University of Chemical Technology,Shenyang 110142,China;Key Laboratory of Optical-Electronics Information Technology Processing,Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110169,China)

机构地区：[1]沈阳化工大学信息工程学院,辽宁沈阳110142 [2]中国科学院沈阳自动化研究所光电信息处理重点实验室,辽宁沈阳110169

出　　处：《微电子学与计算机》2025年第4期114-123,共10页Microelectronics & Computer

基　　金：辽宁省自然科学基金(2023010411-JH3/101)。

摘　　要：在图像语义分割领域,无监督领域自适应技术的发展有效降低了模型对标注数据的依赖,提升了自动驾驶等智能系统的效率和广泛适用性。针对无监督领域自适应技术在新场景泛化能力有限及在稀有类别中分割效果差的问题,文章提出了一种基于强弱一致性的无监督领域自适应语义分割算法。算法首先通过增加特征级别的增强,拓展图像增强空间的维度,改善了只利用图像级增强局限性。其次,采用基于能量分数的伪标签筛选方法,筛选出足够接近当前训练数据的样本赋予伪标签,避免了使用Softmax置信度方法在稀有类别中存在局限性,使模型更新更加稳健。最后,构建结合图像级别增强和特征级别增强的双重一致性框架,更充分的利用一致性训练,进一步提高模型的泛化能力。实验结果证明,提出的方法在GTA5-to-Cityscapes公开数据集中平均交并比指标(mean Intersection over Union,mIoU)可提升至52.6%,较PixMatch算法,性能提升了4.3%。In the field of image semantic segmentation,the development of Unsupervised Domain Adaptation(UDA)techniques has significantly reduced the model's reliance on annotated data,improving the efficiency and broad applicability of intelligent systems,such as autonomous driving.This paper addresses the limited generalization ability of UDA techniques in new domains and poor segmentation performance in rare categories.We propose an unsupervised domain adaptive semantic segmentation algorithm based on strong-weak consistency.The algorithm first extends the image enhancement space through feature-level augmentation,overcoming the limitations of relying solely on image-level augmentation.It then applies an energy score-based pseudo-label filtering method to select samples close to the current training data,avoiding the limitations of softmax confidence in rare categories and making model updates more robust.Finally,we construct a dual consistency framework that combines both image-level and feature-level enhancements to fully leverage consistency training, further improving the model’s generalization ability. Experimental results demonstrate that the proposed method achieves a mean Intersection over Union (mIoU) of 52.6% on the public GTA5-to-Cityscapes dataset, improving by 4.3% over the PixMatch algorithm, showing its potential for both research and practical applications.

关 键 词：语义分割 无监督领域自适应 强弱一致性 伪标签 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]