基于历史分类加权和分级竞争采样的多视角主动学习

发布时间：2018-02-06 04:56

本文关键词： 多视角主动学习分类器集成强化分级竞争采样图像分类　出处：《电子学报》2017年01期 　论文类型：期刊论文

【摘要】：多视角主动学习是一种相比于传统主动学习能够取得更大程度版本空间缩减的技术,已被应用于多种类型的大数据分析中.本文针对现有的多视角主动学习算法在分类假设生成和采样策略中存在的不足分别提出了相应的改进方案.本文将Boosting思想应用到多视角主动学习框架中,通过将历史上各次查询得到的分类假设进行加权式投票来实现每次查询后分类假设的强化;与此同时,还提出了一种自适应的分级竞争采样策略,当分类争议样本规模较大时通过无监督谱聚类获得上述样本的空间分布描述,并在各个聚类中结合样本的分类不确定度和冗余度信息通过二次规划求解以获得可靠的批处理采样.为了证明上述改进的有效性,本文将多视角主动学习应用到图像分类领域中,并通过基于不同图像特征的视角来分别生成相应的分类假设.实验表明,本文提出的两点改进策略不仅均有助于提升多视角主动学习的性能,而且基于上述不同视角随机组合的多视角主动学习方法相比于经典的单视角主动学习算法能够更快地实现收敛并达到较高的场景分类准确性.
[Abstract]:Multi-view active learning is a technology that can reduce version space more than traditional active learning. It has been applied to many types of big data analysis. In this paper, we propose corresponding improvement schemes for the shortcomings of existing active learning algorithms in classification hypothesis generation and sampling strategy. In this paper, Boo is proposed. Sting is applied to the framework of multi-perspective active learning. The classification assumptions obtained from each query in history are weighted to vote to realize the enhancement of classification assumptions after each query. At the same time, an adaptive hierarchical competitive sampling strategy is proposed. When the sample size is large, unsupervised spectral clustering is used to obtain the spatial distribution description of the above samples. In order to prove the effectiveness of the above improvements, the classification uncertainty and redundancy information of the samples are solved by quadratic programming in each cluster to obtain reliable batch sampling. In this paper, multi-view active learning is applied to the field of image classification, and the corresponding classification hypotheses are generated based on different image features. The two improved strategies proposed in this paper are not only helpful to improve the performance of multi-perspective active learning. Moreover, the multi-view active learning method based on the random combination of different angles of view can achieve faster convergence and achieve higher accuracy of scene classification than the classical single-view active learning algorithm.
【作者单位】：宁波工程学院电子与信息工程学院;
【基金】：浙江省自然科学基金(No.LY15F020011,No.LQ15F020004) 浙江省公益技术应用研究计划(No.2016C33255) 宁波市自然科学基金(No.2015A610132)
【分类号】：TP391.41;TP181

【相似文献】