计算机应用研究2024，Vol.41Issue(11)：3337-3342,6.DOI:10.19734/j.issn.1001-3695.2024.02.0070
基于改进双重深度Q网络主动学习语义分割模型
Active learning semantic segmentation model based on improved double deep Q network
摘要
Abstract
This paper proposed an active learning semantic segmentation model named CG_D3QN,based on an improved dual deep Q-network,to address the challenges of acquiring pixel labels and class imbalances in image semantic segmentation tasks.The model used a hybrid network structure that integrates a dueling network architecture with gated recurrent units.This structure alleviated the overestimation of Q-value and efficiently utilized historical state information,thereby improving the ac-curacy and computational efficiency of policy evaluation.On the CamVid and Cityscapes datasets,the model reduced the re-quired sample annotation volume by 65.0％and enhanced the mean intersection over union by approximately 1％to 3％for classes with fewer sample labels.Experimental results show that the model significantly reduces the cost of sample annotations and effectively mitigates class imbalance issues,while being adaptable to different segmentation networks.关键词
深度强化学习/主动学习/图像语义分割/决斗网络/门控循环单元Key words
deep reinforcement learning/active learning/image semantic segmentation/dueling network/gate recurrent unit分类
信息技术与安全科学引用本文复制引用
李林,刘政,南海,张泽崴,魏晔..基于改进双重深度Q网络主动学习语义分割模型[J].计算机应用研究,2024,41(11):3337-3342,6.基金项目
重庆市教育委员会科学技术研究项目(KJQN202101149) （KJQN202101149）
重庆市基础研究与前沿探索专项资助项目(CSTB2022NSCQ-MSX0918,CSTB2022NSCQ-MSX0493) （CSTB2022NSCQ-MSX0918,CSTB2022NSCQ-MSX0493）
重庆理工大学研究生创新资助项目(gzlcx20233251) （gzlcx20233251）
