基于主动迁移学习的图像目标自动标注

Automatic Annotation for Image Objects Based on Active Transfer Learning

针对目标检测中海量图像数据缺乏标注的情况,提出了一种基于主动学习和域适应迁移学习的联合训练策略。为了显著降低标注量,引入了主动学习,利用了不确定策略和无监督聚类算法来采样关键数据进行标注;引入了域适应迁移学习,通过类似的已标注的源域数据集拟合无标注的目标域数据集。将2种方法联合进行数据标注,在速度和精度上均获得更优的效果。在自行收集的数据集上试验结果证明了该方法的有效性。

Aimed at the lack of annotation for object detection in massive image data,a joint training strategy based on the active learning and the domain adaptive transfer learning is proposed.To signifi⁃cantly reduce the amount of annotation,the active learning is introduced,and the uncertain strategy and the unsupervised clustering algorithm are used to sample key data for annotation.The domain adaptive transfer learning is also introduced,and the unlabeled target domain data sets are fitted by the similar annotated source domain data sets.The two methods are combined for data annotation,and the better effect on the speed and the accuracy are gained.Experimental result on the self-collected da⁃ta sets shows the effectiveness of the method.