通信与信息技术Issue(z1)：16-20,5.
一种面向单源域自适应的免训练推理框架
A training-free inference framework for single-source domain adaptation
LIU Dongyin 1WEI Huangsong 1HAN Jianxun1
作者信息
- 1. Neijiang Branch,China Telecom Corporation Limited,Neijiang 641100,China
- 折叠
摘要
Abstract
Single-Source Domain Adaptive(SSDA)object detection aims to utilize a labeled single source domain to train a detection mod-el,further enhancing its generalization ability in different unknown target domains.In this paper,we propose a reasoning enhancement framework without additional training,focusing on strengthening effective features to suppress noise interference,thereby improving the generalization ability of the detection model in the target domain.Specifically,we first analyze pixel features,predicting its probability weight as a foreground target fea-ture and actual activation weight.Based on these two probability weights,we further design a weight-guided Dropout mechanism to dynamically eliminate redundant or interfering components in the feature vector.Finally,by designing an iterative analysis method,we evaluate the contribu-tion of processed features to the attention of query Tokens,ultimately achieving attention fusion and optimization.Experiments on multiple stan-dard datasets prove that this method,under the premise of no training,still demonstrates superior performance and good cross-domain generaliza-tion ability.关键词
单源域自适应/免训练/推理框架Key words
Single-source domain adaptive/Training-free/Inference framework分类
信息技术与安全科学引用本文复制引用
LIU Dongyin,WEI Huangsong,HAN Jianxun..一种面向单源域自适应的免训练推理框架[J].通信与信息技术,2025,(z1):16-20,5.
