基于跨领域主动学习的图像分类方法

Cross-domain active learning algorithm for image classification

针对基于单一领域主动学习的图像分类方法不能利用不同领域图像共同特征导致标记效率低下的问题,提出一种基于跨领域主动学习的图像分类方法。由不同领域图像学习出含共同隐特征的子空间,综合考虑共同特征和领域相关特征,将数据实例引起的模型损失减少量分解到一个共同部分和领域相关部分,从而领域间的共同信息可以编码到模型损失减少的共同部分并用来进行查询。实验结果显示该方法相对于单一模型学习和混合模型学习方法可以减少将近30%的标记工作,并且可以获得更高的精度,表明该方法可以更高效地运用于各种图像分类任务。

To solve the problem of ignoring common information in different domains in traditional active learning image classification,a new multi-domain active learning image classification method was proposed to reduce the labeling effort of image classification.It learned a subspace to represent common features among different images.Considering the common features and domain-specific features,the model loss due to each data instance could be divided into two parts,so that the common information could be queried from the common part.The experimental results show that the new method has some precise increase and has 30％ less labeling efforts than the single model method and mixture model method.The results reveal that the new method can be widely used in all kinds of image classification tasks with higher precise and efficiency.