摘要
跨域目标检测是最近兴起的研究方向,旨在解决训练集到测试集的泛化问题.在已有的方法中利用图像风格转换并在转换后的数据集上训练模型是一个有效的方法,然而这一方法存在不能端到端训练的问题,效率低,流程繁琐.为此,我们提出一种新的基于图像风格迁移的跨域目标检测算法,可以把图像风格迁移和目标检测结合在一起,进行端到端训练,大大简化训练流程,在几个常见数据集上的结果证明了该模型的有效性.
Cross-domain object detection is a new research direction,which aims to solve the problem of generalization from training set to test set.In the existing methods,using image style transfer and train the model on the converted data set is an effective method.However,this method has the problems of not end-to-end training,low efficiency,and tedious process.Therefore,we propose a new cross domain target detection algorithm based on image style migration,which can combine image style migration and target detection to carry out end-to-end training,and greatly simplify the training process.The results on several common datasets show the validity of the model.
作者
吴泽远
朱明
WU Ze-Yuan;ZHU Ming(School of Information Science and Technology,University of Science and Technology of China,Hefei 230026,China)
出处
《计算机系统应用》
2021年第1期194-199,共6页
Computer Systems & Applications
基金
安徽省重点研发计划(201904a05020035)。
关键词
跨域
目标检测
风格迁移
端到端
cross-domain
object detection
style transfer
end-to-end
