| 语义级正则化加持的跨域目标检测方法 |
投稿时间：2025-09-19 修订日期：2025-11-11  点此下载全文 |
| 引用本文： |
| 摘要点击次数: 17 |
| 全文下载次数: 0 |
|
| 基金项目:国家自然科学基金项目：62176107 镇江市重点研发计划项目：GY2023045 |
|
| 中文摘要:域偏移问题是目标检测任务面临的一项重要挑战。作为无监督域自适应中的一项基准方法，对抗特征对齐已广泛用于增强检测器的域泛化能力。然而，现有方法通常聚焦于图像内部特征的一致性，忽视了针对核心目标的语义性监督。对此，本文提出一种语义级正则化加持的域自适应目标检测方法（SRDA）。针对已标注源域数据，通过在主干网络末端添加多标签分类器来矫正目标实例的语义信息。针对无标注目标域数据，通过在检测头尾端附加熵最小化损失，来强化检测器对目标区域的鉴别性。SRDA以单阶段YOLOv9-S为基准检测器，在多个跨域场景下取得了优异性能。 |
| 中文关键词:域偏移 对抗特征学习 语义级正则化 单阶段检测器 |
| |
| Semantic Regularization Enhanced Domain Adaptive Object Detection |
|
|
| Abstract:Domain shift poses a significant challenge in object detection tasks. As a benchmark method in unsupervised domain adaptation, adversarial feature alignment has been widely employed to enhance the domain generalization capability of detectors. However, existing methods often focus on achieving consistency within image features while neglecting semantic supervision for core target regions. To address this issue, this paper proposes a Semantic Regularization Enhanced Domain Adaptive Object Detection (SRDA) method. For labeled source domain data, a multi-label classifier is added at the end of the backbone network to rectify the semantic information of object instances. For unlabeled target domain data, an entropy-minimization loss is applied at the end of the detection head to strengthen the detector’s discriminative ability for target regions. SRDA, built upon the single-stage YOLOv9-S detector, achieves superior performance across multiple cross-domain scenarios. |
| keywords:Domain shift adversarial feature learning semantic regularization single-stage detector |
| 查看全文 查看/发表评论 下载pdf阅读器 |
次