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 onimproved double deep Q network

作　　者：李林 刘政[1] 南海 张泽崴 魏晔 Li Lin;Liu Zheng;Nan Hai;Zhang Zewei;Wei Ye(College of Computer Science&Engineering,Chongqing University of Technology,Chongqing 400054,China)

机构地区：[1]重庆理工大学计算机科学与工程学院,重庆400054

出　　处：《计算机应用研究》2024年第11期3337-3342,共6页Application Research of Computers

基　　金：重庆市教育委员会科学技术研究项目(KJQN202101149);重庆市基础研究与前沿探索专项资助项目(CSTB2022NSCQ-MSX0918,CSTB2022NSCQ-MSX0493);重庆理工大学研究生创新资助项目(gzlcx20233251)。

摘　　要：针对在图像语义分割任务中获取像素标签困难和分割数据集类别不平衡的问题,提出了一种基于改进双重深度Q网络的主动学习语义分割模型CG_D3QN。引入了一种结合决斗网络结构以及门控循环单元的混合网络结构,通过减轻Q值过估计问题和有效地利用历史状态信息,提高了策略评估的准确性和计算效率。在CamVid和Cityscapes数据集上,该模型相较于基线方法,所需的样本标注量减少了65.0%,同时对于少样本标签的类别,其平均交并比提升了约1%~3%。实验结果表明,该模型能够显著减少样本标注成本并有效地缓解了类别不平衡问题,且对于不同的分割网络也具有适用性。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 accuracy and computational efficiency of policy evaluation.On the CamVid and Cityscapes datasets,the model reduced the required 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 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]