基于环境异质信息的机器觉察与仿生知觉方法研究

Machine awareness and bio-inspired perception are the important research field and key technology of machine environment perception. The main task of machine awareness is to achieve the awareness of environment information in machine by simulating the biological awareness and perception characteristic in processing heterogeneous information. However, current approaches of realizing machine awareness lack of effective model of human awareness characteristic as well as the perceptual processing protocol. Motivated by this, we propose and study the artificial awareness modeling and perceptual processing procedure for heterogeneous information based machine awareness and bio-inspired perception. Two outcomes are expected, which are the new concepts and effective approaches for machine awareness and perception, including the awareness models of auditory and visual channels, bio-inspired perception models of heterogeneous information and perceptual processing framework. Particularly, for the auditory awareness, the biological inhibition of return effect modeling is focused on, and the multi-scale saliency feature fusion method is studied. Meanwhile, for the visual awareness, the objectness measuring method based on reconstruction error is concentrated and bio-inspired visual awareness mechanism is researched by applying visual saliency and informatics saliency. New framework of perceptual processing procedure is expected to simulate the human perceptual characteristic, and to concatenate heterogeneous information, environmental object and scene together compactly, so that multiple perception tasks can be accomplished by machine. The achievements of the project will bridge the gap between the awareness and perception of environment, and will be very useful for the further research and application of artificial intelligence.

机器觉察和仿生知觉是机器环境感知研究的重要方向和核心内容，是通过模拟人类处理异质信息时的生物觉察与知觉特性，实现机器对环境信息的智能感知。然而现有的环境信息觉察方法缺乏对人类觉察特性的有效建模，更缺乏异质信息仿生知觉处理的可行方法。因此在本项目中，拟对听觉返回抑制效应进行仿生建模，研究并设计多尺度显著性特征的计算、校验与融合方法，提出模仿人类听觉觉察特性的机器觉察方法。同时，拟对前景目标的物体性检测进行研究，设计基于物体性的信息显著性表达与检测方法，对知觉显著的前景目标进行仿生觉察。随后，模仿人类对异质信息处理的知觉一致性，研究并提出基于信息概率形式和结构性显著的仿生知觉方法，以期提高机器环境感知的适应能力和智能度。通过本项目的研究，从智能信息处理角度对实现机器觉察和仿生知觉提供系统的解决方案，为丰富和发展异质信息觉察与知觉处理做出积极的贡献，为人类觉察与知觉机制建模进行有益的尝试。

在模拟人类对环境进行感知时的生物觉察特性基础上，本项目针对视听觉信息的机器觉察和仿生知觉方法进行了研究。首先，研究了多尺度显著性特征提取方法，对听觉返回抑制效应的机制、模式和参数进行了分析，建立了听觉显著性检测模型，构建了面向不同背景的环境声音信号数据库，提出了基于多尺度特征融合的环境声音信号感知方法。其次，研究了基于信息显著性的前景显著目标检测方法，从特征表达、运动建模两个方面对视觉觉察的计算化机制进行了研究，建立了面向环境弱小视觉目标的图像数据库，提出了面向弱小前景目标的检测方法，并提出了基于深度特征抽取的视觉觉察方法。最后，研究了基于环境异质信息的仿生建模方法与知觉处理流程，建立了面向环境物体的仿生知觉模型和面向环境场景的仿生知觉模型，提出了基于知觉负属性建模的结构化显著性表征机制，并设计了基于视听觉信息的仿生知觉处理框架。本项目的研究为自主无人系统的智能环境感知任务提供了基础性技术支持，对人机共融社会的发展具有重要理论意义和实用价值。