基于主动学习的人脸标注研究

Research on Face Tagging Based on Active Learning

在大数据时代,图片数量非常巨大,但是具有标签的图片非常少。在学习和研究中,常常需要分类标注图片,而大部分图片都是与人脸相关的,因此人脸标注成为了一种进行图片分类标注的有效方法,但人工标注的成本较大。针对有标签图片数量较少以及人工标注成本较大的问题,提出了在主动学习算法的基础上建立计算人脸类标签后验分布的判别模型的方法。该方法基于马尔可夫随机场和高斯过程,考虑到了样本位置、特征的客观联系,在样本之间加入了匹配约束和非匹配约束,匹配约束表示样本之间具有相同的类标签,非匹配约束表示样本之间具有不同的类标签。实验结果表明,根据判别模型得到的类标签后验分布选择样本进行人工标注,大大提高了分类器的精确度。

In the era of big data,tremendous images are available,whereas images with tags are sparse relatively.For the purpose of learning and research,it’s necessary to classify and annotate images,and most images are relevant to faces,consequently face tagging is an effective tool to annotate images.However,the cost of manual annotation is high.Aiming at solving the problems of lacking tagged images and high manual annotation cost,a discriminative model based on the active learning inducing the posterior distribution over labels was proposed.The discriminative model is based on markov random field(MRF)and gaussian process(GP),and considers the objective connections between the positions and features of samples with the addition of match constraint and non-match constraint between samples.Match constraint means that samples have the same label,while non-match constraint means that samples have different labels.Experimental results indicate that choosing samples for manual annotation according to the posterior distribution over labels induced by the discriminative model can greatly improve the classification accuracy.