基于多视角二维主动学习的多标签分类

Multi-label classification based on multi-view two-dimensional active learning

针对多标签图像分类问题的特点，提出了一种多视角二维主动学习（MV-2DAL）算法，以通过多视角学习与主动学习的有机结合，深入挖掘样本、标签、视角三个维度上的相关性和冗余性。此算法以样本．标签对作为基本标注单位，在每个视角内，利用二维主动学习的方法计算样本、标签维度上的不确定度；在不同视角间，通过多视角融合的方法计算跨视角的不确定度；最终，将视角内不确定度与视角间不确定度进行融合得到总不确定度，并以此衡量样本-标签对的标注价值。将MV-2DAL算法应用到图像内容理解的一个重要领域——多标签图像分类中，显著提高了信息标注的针对性，不仅有效降低了信息冗余度，同时也大幅减少了数据标注量。

This paper presents the multi-view two-dimensional active learning （MV-2DAL） algorithm for multi-label image classification so as to thoroughly explore the redundancies along the dimensions of sample, label and view, by the organic integration of the active learning with the multi-view learning. Taking a sample-label pair as the basic labeling unit, the algorithm calculates the uncertainties from the dimensions of sample and label within each view u- sing the two-dimensional active learning, and captures the uncertainties over different views based on the muhi-view fusion. The overall uncertainty along the three dimensions is obtained to detect the most informative sample-label pairs. The experiments on the real-world multi-label image classification demonstrate that the proposed MV-2DAL algorithm is effective for redundancy reduction, and thus greatly alleviates the burden on human labeling.