基于特征不变约束的恶劣场景语义分割
集成电路应用
刘溪源,杨尚宗,赵宇
江苏科技大学计算机学院
摘要: 针对恶劣天气和夜间场景中语义分割模型易受光照、颜色和噪声扰动影响的问题,该文提出了一种基于特征不变性约束的无监督领域自适应方法。该方法以CoDA框架为基础,引入浅层风格迁移构造语义一致、外观不同的双视图,并在编码器特征空间约束原始视图与风格化视图的一致性,从而减少外观变化对语义表示的干扰。实验表明,该文方法在多个恶劣场景数据集上均优于CoDA基线,并在ACDCAll数据集上mIoU达到69.10%,验证了该方法的有效性和泛化能力。
中图分类号:TP391.4文献标识码:ADOI:10.19339/j.issn.1674-2583.2026.08.006
中文引用格式:刘溪源,杨尚宗,赵宇. 基于特征不变约束的恶劣场景语义分割[J].集成电路应用,2026,43(8):28-31.
英文引用格式:Liu Xiyuan, Yang Shangzong, Zhao Yu. Adverse scene semantic segmentation based on feature invariance constraints[J].Application of IC,2026,43(8):28-31.
中文引用格式:刘溪源,杨尚宗,赵宇. 基于特征不变约束的恶劣场景语义分割[J].集成电路应用,2026,43(8):28-31.
英文引用格式:Liu Xiyuan, Yang Shangzong, Zhao Yu. Adverse scene semantic segmentation based on feature invariance constraints[J].Application of IC,2026,43(8):28-31.
Adverse scene semantic segmentation based on feature invariance constraints
Liu Xiyuan, Yang Shangzong, Zhao Yu
School of Computer Science,Jiangsu University of Science and Technology
Abstract: To improve semantic segmentation in adverse weather and nighttime scenes, where the model is easily disturbed by illumination, color and noise perturbations, this paper proposes an unsupervised domain adaptation method based on feature invariance constraints. Built on CoDA, the method uses shallow style transfer to construct dual views that are semantically consistent but visually different, and constrains the consistency between the original view and the stylized view in the encoder feature space to reduce the impact of appearance variations. Experiments show that the proposed method outperforms the CoDA baseline on multiple adverse-scene datasets, achieving 69.10% mIoU on ACDC-All, which verifies its effectiveness and generalization ability.
Key words : adverse scene parsing; unsupervised domain adaptation; feature invariance learning
引言
无监督领域自适应语义分割旨在利用带标注源域数据和无标注目标域数据,使模型在目标场景中保持稳定预测。自动驾驶任务中,模型多在白天清晰场景上训练,而实际部署常面临夜间、雾天、雨天和雪天等复杂环境。亮度不足、颜色偏移和噪声等因素会造成明显域差异,影响道路、天空、车辆及远距离目标的识别。现有方法[1-2]主要通过对抗学习、自训练和风格迁移缩小域差异。但自训练依赖伪标签质量,输出层一致性约束易受低质量伪标签影响,单纯风格迁移也难以保证中间语义特征稳定。链式域自适应(ChainofDomain Adaptation,CoDA)[3]框架通过链式域适应和视觉提示机制实现恶劣场景渐进迁移,但对风格扰动下的特征稳定性建模仍不足。为此,本文在CoDA框架中引入特征不变性约束,利用跨风格双视图学习稳定的编码器表示。主要贡献为:构造轻量级跨风格双视图模拟外观扰动;在编码器空间引入特征不变损失,减少噪声监督影响;在多个恶劣场景数据集上验证方法的适应性和泛化能力。
本文详细内容请下载:
https://www.chinaaet.com/resource/share/2000007320
投稿邮箱:ICAPP@belling.com.cn
作者信息:
刘溪源,杨尚宗,赵宇
(江苏科技大学计算机学院,江苏镇江212003)

此内容为AET网站原创,未经授权禁止转载。
