基于特征学习和标记关联的多标记学习算法研究

Multi-label learning, which considers the case of an object related to multiple　labels, attracts much attention in recent years. However, the performance of　multi-label learning is still not satisfied due to the lack of the effective abstract features. Moreover, the multi-label learning also suffers from the high-dimensional feature space and the difficulty of modeling the higher-order correlation among labels. Focusing on the multi-label automatic image annotation, this project studies the following problems. Firstly, considering the problem that the multi-label learning uses the hand-crafted features as the input, an automatic feature learning algorithm is proposed based on deep learning. Secondly, considering the difficult of modeling the correlation among labels, a modeling method based on conditional random field and semantic calculation is proposed to model and quantize the correlation among labels. Finally, considering the effect of label correlation, the multi-label classification algorithm constrained by label dependency is proposed. Different from the traditional transductive inference method, this proposal adopts the strategy of ‘divide and rule’, which resolves the three complex problems of feature representation, label correlation modeling and multi-label classification respectively. It enriches and extends the theories and methods of machine learning and has important guiding significance for automatic image annotation.

多标记学习考虑一个对象与多个类别标记相联系的情况，是当前国际机器学习领域研究的热点问题之一。然而，由于多标记学习依然使用底层特征，其学习的准确率有待提高。此外，特征空间维度过高和标记之间高阶依赖关系建模困难也成为制约该领域研究的瓶颈。本项目以图像自动标注为背景，研究以下问题：1）针对底层特征抽象程度不高影响分类精度的问题，拟通过深度学习方法自动学习能刻画数据本质的特征并且降低特征空间维度；2）针对标记高阶关联依赖难建模问题，拟通过概率图模型和语义计算对标记之间依赖关系建模，实现标记关联关系的量化描述；3）针对标记关联关系影响多标记分类结果的问题，拟通过最优化方法训练标记关系对多标记学习的约束参数，提高多标记分类性能。不同于传统的直推式学习方法，采取将特征学习、标记关联性分析和多标记分类三个复杂问题分而治之的思路，丰富和拓展了机器学习理论和方法，对于解决图像自动标注问题具有重要的指导意义。

本项目针对多标记学习中存在的特征空间维度过高和标记之间高阶依赖关系建模难等问题，以降低特征空间和标记空间的复杂度，提升多标记学习算法的精度为目标，开展多标记学习算法的研究，主要取得如下成果： .（1）借鉴粒计算思想，针对多标记数据的特点，结合标记之间的相关性特征，提出不同层次下知识粒的相似性度量方法和粒化方法。然后在已获取的粒化结构的基础上，建立多标记分类模型。.（2）针对多标记学习中存在的特征抽象层次不高而维度过高的问题，利用包含多个隐含层的深度神经网络直接从原始的输入中学习并构建多层的分级特征，形成更加抽象的高层表示，力求以最少和最有效的特征来表达原始多标记数据。.(3）针对多标记样本的标记之间存在关联依赖且高阶依赖关系难学习问题，借助贝叶斯网络模型捕捉变量之间的高阶依赖关系，构建标记关系分析算法，对标记结构关系进行量化描述，并在该结构关系约束下设计多标记分类算法。.(4)探索了多标记学习理论和方法在图像自助标注、文本信息处理等领域的应用。通过深度神经网络提取多标记图像或文本数据特征，并设计了数字图像自动标注系统和文本联机分析处理系统。.（5）基于粒计算思想，改进数字图像水印分割模型，即将水印图像进行多粒度分解，得到信息粒后，再进行压缩感知以获取测量值。经过多粒度分解处理的水印图像与原图像相似度极高，不仅鲁棒性好而且不可见性较强。