属性知识引导的自适应视觉感知与结构理解研究进展

Progress in Attribution-Guided Adaptive Visual Perception and Structure Understanding

机器通过自适应感知从环境中提取人类可理解的信息,从而在开放场景中构建类人智能.因属性知识具有类别无关的特性,以其为基础构建的感知模型与算法引起广泛关注.文中首先介绍属性知识引导的自适应视觉感知与结构理解的相关任务,分析其适用场景.然后,总结四个关键方面的代表性工作.1)视觉基元属性知识提取方法,涵盖底层几何属性和高层认知属性;2)属性知识引导的弱监督视觉感知,包括数据标签受限情况下的弱监督学习与无监督学习;3)图像无监督自主学习,包括自监督对比学习和无监督共性学习;4)场景图像结构化表示和理解及其应用.最后,讨论目前研究存在的不足,分析有价值的潜在研究方向,如大规模多属性基准数据集构建、多模态属性知识提取、属性知识感知模型场景泛化、轻量级属性知识引导的模型开发、场景图像表示的实际应用等.

Machines extract human-understandable information from the environment via adaptive perception to build intelligent system in open-world scenarios.Derived from the class-agnostic characteristics of attribute knowledge,attribution-guided perception methods and models are established and widely studied.In this paper,the tasks involved in attribution-guided adaptive visual perception and structure understanding are firstly introduced,and their applicable scenarios are analyzed.The representative research on four key aspects is summarized.Basic visual attribute knowledge extraction methods cover low-level geometric attributes and high-level cognitive attributes.Attribute knowledge-guided weakly-supervised visual perception includes weakly supervised learning and unsupervised learning under data label restrictions.Image self-supervised learning covers self-supervise contrastive learning and unsupervised commonality learning.Structured representation and understanding of scene images and their applications are introduced as well.Finally,challenges and potential research directions are discussed,such as the construction of large-scale benchmark datasets with multiple attributes,multi-modal attribute knowledge extraction,scene generalization of attribute knowledge perception models,the development of lightweight attribute knowledge-guided models and the practical applications of scene image representation.