可扩展零样本图像分类中异质结构对标记迁移的建模推理研究

Unlike support vector machine and deep learning which need training samples, zero-shot image classification can complete image recognition with the aid of richly side information transfer. Compared with the traditional learning method of sharing text or attribute zero-shot classification, cross-domain label transfer zero-shot classification between side information and image has the more direct advantage because it can construct transfer function or label propagation by exploring the relationship of cross-domain data structure. However, the scalable data lead to data multi-structure changing, and make the poor adaptivity for following the variety data on the fixed transfer function or label propagation. This project builds heterogeneous structure modeling and reasoning for label transfer with mining heterogeneous structure relationship of cross-domain multi-structure by the adaptively dynamic process. Heterogeneous structure on the core regulation of label can be found out; the effect rules of label on heterogeneous structure relation can be revealed; further, label transfer of causal reasoning, evolution laws, and coupling mechanism of heterogeneous structure relationship can be discovered; finally, the adaptive computation model of label transfer based on heterogeneous structure can be constructed, and the scalable, general and feasible method can be obtained for zero-shot image classification and understanding. This project can enrich heterogeneous structure cognition and label transfer theory in data science, and provide the theoretic and technological base for large-scale image recognition and understanding.

不同于需要训练样本的支持向量机和深度学习，零样本图像分类能借助于丰富的辅助信息迁移完成图像识别。相比传统通过学习共享文本或属性的零样本分类方法，在辅助信息和图像数据中的跨域标记迁移零样本分类，因为其能挖掘数据分布结构关系建立迁移函数或标记传播，表现出更为直接的优势。然而数据的扩展导致数据间的多重结构关系在不断变化，使得固定的迁移函数或标记传播并不具有数据扩展变化的适应性。本项目挖掘跨域多重结构的异质结构关系，通过自适应动态过程，建立异质结构对标记传播和迁移推理模型。研究发现异质结构对标记迁移具有的关键调控作用；挖掘标记对异质结构关系影响规律；进而揭示异质结构与标记迁移的因果关系、演变规则及耦合机制；最终建立自适应的异质结构和标记迁移的计算模型，获得可扩展零样本图像理解分类的普适和可行的方法。本项目可丰富数据科学异质结构认知与标记迁移理论体系，为大规模零样本图像识别理解提供理论和技术基础。

不同于传统通过学习共享文本或属性的零样本分类方法，在辅助信息和图像数据中的跨域标记迁移零样本分类，因为其能挖掘数据分布结构关系建立迁移函数或标记传播，表现出更为直接的优势。因为数据的扩展导致数据间的多重结构关系在不断变化，使得固定的迁移函数或标记传播并不具有数据扩展变化的适应性。本项目挖掘跨域多重结构的异质结构关系，通过自适应动态过程，建立异质结构对标记传播和迁移推理模型。主要的研究有以下三方面：①研究异质结构相互作用机理，随数据的不确定性和动态扩展性的异质结构及关系；②研究标记和异质结构因果关系推理学习模型，研究贝叶斯网络推理及深度学习在标记中对异质结构及关系影响机制；③研究标记传播控制和异质结构关系的动态作用机制，提出标记与异质结构及关系的动态交互方法。本项目可丰富数据科学异质结构认知与标记迁移理论体系，为大规模零样本图像识别理解提供理论和技术基础。