基于主动学习的唐卡主尊标注研究

Research on Thangka Yidam annotation based on active learning

唐卡是藏文化中一种特色绘画,有较高的学术价值。为降低唐卡主尊检测任务中数据标注的人力成本,文中采用主动学习流程训练目标检测模型,并针对唐卡主尊分布特点提出最大框选法和最大框不确定性方法,用以优先选取未标注样本中对神经网络最有益的样本。采用Faster R-CNN目标检测模型进行唐卡主尊主动学习实验,结果表明:所提出的最大框不确定性方法优于随机采样主动学习方法,仅400张训练数据即可达到98.19%的平均准确率(mAP),与全监督下1 249张数据训练的模型结果(98.17%)接近;在500张数据时mAP可达到最高,为98.31%。所提最大框不确定性采样法可高效挑选出高信息量唐卡主尊数据,不但可以降低训练所需数据量,减少网络训练时间,而且能够减少低信息量数据对模型的影响,对模型的性能具有显著提升效果。

Thangka is a characteristic painting in Tibetan culture and has high academic value.In order to reduce the labor cost of data annotation in the Thangka Yidam detection task,the active learning is used to train the object detection model,and a largest box selection method and the largest box uncertainty are proposed according to the characteristics of the Thangka Yidam,so as to preferentially select the samples that are most beneficial to the neural network among the unlabeled samples.Experiments with Thangka Yidam active learning on the Faster R⁃CNN object detection model show that the proposed maximum box uncertainty method is comprehensively higher than the randomized sampling active learning method,and can reach 98.19%mAP with only 400 training data,which is close to the results of 1249 data training under full supervision(98.17%).At 500 data sheets,mAP can reach its highest level,reaching 98.31%.The proposed maximum box uncertainty sampling method can efficiently select high⁃information Thangka Yidam data,which not only reduces the amount of data required for training and reduces network training time,but also reduces the impact of low information data on the model,significantly improving the performance of the model.