全局-实例特征对齐域适应检测方法及系统设计
Global-instance feature alignment domain adaptation detection method and system design
刘源 1娄亚鑫 1张平 2杨一帆 1李亚伟 1伍凌帆 1张弘1
作者信息
- 1. 北京航空航天大学宇航学院,北京 102206
- 2. 93129 部队,北京 100036
- 折叠
摘要
在实际应用检测模型时,由于真实场景和训练数据集间的差异,检测算法的效果受到较大影响.为了在目标场景中获得更好的检测效果,通常需要采集、标注数据后训练,不仅成本高昂且流程复杂.提出基于注意力机制的全局-实例域适应检测算法与系统,仅需采集部分真实场景数据即可进行迁移学习,实现模型快速训练和边缘端-云端结合的远程部署.该域适应检测算法中,基于注意力机制的全局特征对抗学习算法可减弱背景特征在迁移学习中的负作用;基于字典学习的实例级特征对齐方法,对实例级特征进行高精度对齐.经过实验对比,本文的方法达到了接近SOTA(state-of-the-art)的水平,并通过消融实验定量地证明了本方法对于域适应检测效果的提升.本文将提出的域适应检测技术与具有数据传输链路的边缘端系统结合,在实际场景中使检测效果提升近10个点.
Abstract
When actually applying the detection model,due to the difference between the real scene and the training data set,the effect of the detection algorithm is greatly affected.In order to obtain the better detection effect in the target scene,it is usually necessary to collect and label data and then train,which is not only costly but also complicated.The proposed global-instance domain adaptation detection algorithm and system based on the attention mechanism only needed to collect part of the real scene data to perform transfer learning,realizing rapid model training and remote deployment of edge-cloud integration.In this domain adaptation detection algorithm,the global feature adversarial learning algorithm based on the attention mechanism could reduce the negative effect of background features in transfer learning;the instance-level feature alignment method based on dictionary learning could align instance-level features with high precision.After experimental comparison,the proposed method reached a level close to SOTA(state-of-the-art),and the ablation experiment was quantitatively proved the improvement of the domain adaptation detection effect of this method.The proposed domain adaptation detection technology is combined with an edge system with data transmission links,improving the detection effect by nearly 10 points in actual scenarios.
关键词
域适应检测/实例级/注意力/边缘端设备Key words
domain adaptation detection/instance level/attention/edge device引用本文复制引用
出版年
2024
NETL
NSTL
万方数据