动态视觉认知子空间构建与应用

Dynamic visual category hierarchy plays a vital role during human cognition. It is also the truth that the dynamic visual category hierarchy simulated by computer can improve the performance for lots of methods in the field of computer vision. In this project, the goal is to construct the “visual cognition sub-space”, and consequently the dynamic category hierarchy, by combing the dynamic cognition and sub-space theory. To achieve this goal, we will do a series of researches which will be detailed in the following. Firstly, we will propose a novel real-value representation of semantic attributes and the model which can simultaneously learn and predict multiply attributes, by employing sparse representation, sub-space clustering and multi-label learning. Secondly, we will propose a novel distance-dependent supervised hierarchical topic model based on the object description by employing real-valued semantic attribute, nonparametric probability model, continuous probability distribution, and reconfiguration of dependent relationship. Through this novel topic model, the dynamic category hierarchy can be obtained, and at the same time, we will construct the probability sub-space in the semantic attributes space according to category hierarchy, and form a novel representation of dynamic “visual cognition sub-space” with respect to the category hierarchy. Finally, based on the dynamic “visual cognition sub-space”, the accuracy of object recognition and discovery can also be improved, and the relationship of probability distribution of semantic attributes among seen categories and unseen categories can be formed according the probability sub-space, based on which we will research how to extract the shared pattern and distinctive pattern among categories, and improve efficiently the performance of zero-shot learning.

动态视觉类别层次结构在人类认知过程中起重要作用。同样，计算机模拟人类认知过程建立动态视觉类别层次结构并构建动态“视觉认知子空间”能够极大提升许多计算机视觉方法的性能。本项目从动态认知角度出发，结合子空间理论，构建动态“视觉认知子空间”表达。主要研究内容包括：借鉴稀疏表达、子空间聚类和多标签多样本学习等方法，研究大规模实值语义属性的表示方法，及学习和预测多语义属性的模型；通过语义属性描述物体，引入非参数化概率模型方法，采用连续概率分布建模，重构节点依赖关系，研究新的基于距离依赖和连续分布的有监督层次化主题模型，构建动态层次化类别结构，并创造性地使用语义属性概率子空间表示，形成动态“视觉认知子空间”。最终，应用动态“视觉认知子空间”提升对象识别与发现的准确性；并且建立已见类别和未见类别之间关于语义属性的概率分布的联系，研究类间的共享模式和特定模式的提取，有效提高零样本学习性能。

动态视觉类别层次结构在许多视觉任务中都有非常重要的作用。本项目从如何有效构建类别层次结构和如何应用类别层次结构方面展开研究，完成了无参数概率主题模型的研究、完成了语义属性表征的研究、完成了基于语义特征的零/少样本的半/弱监督深度学习的语义分割应用研究、完成了基于多模态信息融合的环境三维感知研究、完成了基于多终端协作的环境三维感知研究。提出了新的语义属性学习方法、新的层次化非参数概率模型的语义物体层级式表示，基于语义属性和层次化结构表示，实现了弱监督和半监督的语义分割方法，语义SLAM系统，以及多模态融合多终端协作的环境感知系统。项目研究成果提高了领域内对动态视觉类别层次结构的认知，以及在环境感知方面的应用方法，为智能移动终端在复杂环境中进行自主智能探索提供理论基础。