基于知识迁移的有限样本模式分类研究

In many practical application occasions, labeling samples for all classes is time-consuming, laborious and even impossible. Therefore, the pattern classification problem under the learning scene of limited training samples (single-sample or zero-sample) will be researched using knowledge transfer methods in the project. The main contents in our research include the following aspects. Single-sample face recognition based on feature transfer is researched by viewing the samples contained in the face database as transfer sources, which have the same or similar macro facial appearance features as or with the single target training or testing sample. In order to take full advantage of prior knowledge related with attributes under the zero-sample classification scene, two kinds of methods are proposed. Firstly, the intrinsic relation between attributes is mined using structured learning methods and then a multi-attribute joint prediction model is built. Secondly, the intrinsic relation between attributes and low-level features is integrated into the classifier design by borrowing the idea of feature selection existed in multi-task learning. When relative attributes are applied to zero-sample classification problems, all seen and unseen images should satisfy specific distributions. In order to overcome the limitation of relative attributes, the relative attribute is combined with decision tree and then a random forest is built to realize the image classification under zero-sample learning scene. The research fruits not only can enrich the present theory and methods of machine learning, but also can be widely extended to many related fields of pattern classification. Therefore, the research has important theoretical significance and practical value.

在许多实际应用场合，对所有类别的样本进行人工标注是一件费时、费力，甚至是不可能的事情。为此，拟采用知识迁移方法，针对有限训练样本（单样本、零样本）学习场景下的模式分类问题展开研究。内容包括：将人脸数据库中具有与目标训练或测试样本相同或相近宏观特征的样本视为迁移源，研究基于特征迁移的单样本人脸识别；为在零样本学习场景下充分利用与属性相关的先验知识，首先，采用结构化学习方法挖掘属性与属性间存在的内在联系，构建多属性联合预测模型。其次，借鉴多任务学习中特征选择的思路，将属性与底层特征间的联系融入进分类器的设计之中；为克服相对属性需要假定所有未见类图像和可见类图像均服从特定分布的局限性，将相对属性与决策树相结合，进而构造随机森林，实现零样本学习场景下的图像分类。研究成果不仅可以丰富现有的机器学习理论和方法，而且能够广泛推广应用到模式分类的诸多相关领域，具有重大理论意义和实用价值。

采用知识迁移方法，针对有限训练样本学习场景下的模式分类问题展开研究。将人脸数据库中具有与目标训练或测试样本相同或相近宏观特征的样本视为迁移源，提出一种基于特征迁移的单样本人脸识别方法。通过为不同的属性赋予一个相应的权重来区分不同的属性对于决策分类的不同贡献程度，提出一种基于关联概率的加权IAP模型。针对属性学习中存在的所有特征与属性被同等对待，底层特征与属性、属性与属性之间的先验知识被忽略的问题，提出一种基于属性关系图正则化特征选择的零样本分类方法以及基于属性与类别相关先验知识的属性学习模型。针对单纯语义属性描述类别的有限性，提出一种基于混合属性的零样本图像分类模型。鉴于深度学习能够从原始无标签图像中自动提取到更具表达力的特征，分别提出一种基于深度特征提取和深度加权属性预测的零样本学习方法。利用领域适应技术，提出一种基于自适应多核校准的多源域属性适应学习。为克服相对属性需要假定所有未见类图像和可见类图像均服从特定分布的局限性，提出一种基于相对属性的随机森林零样本图像分类方法。针对迁移学习中存在的负迁移、欠适配和欠拟合问题，为学习有效特征并最大程度减小领域间的分布差异，分别提出基于自动编码器的（异构）领域适应网络以及基于超图正则化降噪自动编码器的领域适应网络。为设计更为有效的图像特征提取及分类模型，项目组提出了多种深度学习及维数约简方法，包括：基于金字塔结构的深度卷积网络、多路径集成卷积网络、权重共享多级多尺度集成卷积网络、非负稀疏半监督维数约简、基于成对约束判别分析和非负稀疏散度的降维、基于张量近邻图和补丁校准的降维、基于监督超图和样本扩充CNN的降维等。获教育部自然科学二等奖1项，授权发明专利8件，出版专著4部，培养研究生14名，发表/录用论文40篇（32篇SCI，8篇Ei）。