【作者】 王悦；

【导师】 梁继民；

【作者基本信息】 西安电子科技大学 ， 模式识别与智能系统， 2021， 博士

【摘要】 视觉是人和动物感知外界信息的重要途径,对机体的生存和发展具有重要意义。人类视觉的研究也对计算机视觉的发展有着重要的理论价值。在计算机视觉研究中,了解人类视觉的工作机制是实现独立和自动的图像处理这一目标的重要一步,也是该研究领域的重点和难点之一。然而,目前的计算机视觉方法大多较好的模拟了人类视觉的低层感知过程,并没有对同样重要的人类视觉高级认知过程进行很好的体现,这主要归因于人们对高级视觉处理机制了解的不够充分。视觉信息处理是一个复杂的过程,受多种自下而上刺激驱动因素和自上而下目的导向因素的影响,具有个体特异性,且在不同视觉领域之间具有相似性。基于特定视觉领域的研究是探索该领域特异性视觉机制的有效途径,也能为不同领域间视觉处理机制共性的研究提供重要的理论依据。在特定视觉领域,具有专家化视觉能力的判读员与外行相比具有更高的判断准确率、更快的决策速度、更稳定的行为表现和更强的抗干扰能力。因此,基于专家化视觉能力的研究能够获得更加可靠且更有实际应用价值的结果。专家化视觉能力与广泛的神经系统的活动有关,它不仅涉及视觉系统,还延伸到整个大脑皮层的多个高级功能区,但前人的研究通常是从脑区激活的增加、减少和功能重组方面的变化描述经验对任务执行过程中大脑功能表征的影响,并未从全脑功能交互的角度对其进行研究。大脑神经交互模式的全脑分析需要合理的脑功能区划分,该划分方案具有很强的被试特异性和视觉领域特异性,是该研究的难点之一。此外,对全脑进行功能连接分析时,不可避免的会受到虚假功能连接的影响,如何选择合理的分析方案是该研究的另一个难点。前人对视觉处理中自上而下和自下而上机制的研究通常使用合成或经过人为调整的视觉刺激,这类视觉刺激虽然可以给任务带来可控的变化或难度,使人们可以对关于这些因素的神经编码进行研究,但这类视觉刺激无法体现或保持真实世界中的语义效应而不可避免的剥离了原生背景中一些未知但重要的信息。本文基于原生场景中的车辆检测任务,对真实世界中的目标检测专家化视觉机制进行研究。真实世界场景复杂多样,包含大量信息,但专家化视觉机制的研究需要基于稳定、可靠的脑响应进行,其刺激图像集的构建和实验范式的设计具有很强的任务相关性。因此实验中要在保留有用信息的同时,尽可能的避免干扰因素的影响。针对以上问题,本文基于影像医师模型和原生场景中的车辆检测任务,分别对专家化视觉能力获取所涉及的中枢神经机制和原生场景中目标检测专家化视觉机制的行为和神经表现进行了研究。本文主要工作和研究成果概括如下:(1)基于放射影像医师模型,利用功能磁共振成像和功能连接分析方法探索了放射影像视觉经验对脑区间静息态功能连接可塑性的影响。研究结果表明,放射影像视觉经验可能伴随着感知大脑功能区之间、认知大脑功能区之间较高的功能同步性,认知脑功能之间较高的功能耦合以及感知和认知脑功能之间的功能解耦。本研究弥补了放射视觉经验对脑功能区之间自发连接模式研究的空白,可能为理解现实世界视觉识别专业知识形成的中枢神经机制提供新的思路。(2)针对自然场景中的车辆目标检测任务,设计行为学实验和脑电实验,研究图像特征对被试行为和神经表现的影响,进而研究适用于研究车辆目标检测这项专家化视觉能力的刺激图像集的构建方案和实验范式的设计策略。研究发现,目标与背景间的色调差异对被试目标检测准确率的影响最为显著,刺激图像的场景内容对视觉目标的处理有着重要影响,是视觉目标处理神经机制研究中需要重点考虑的因素。对包含复杂多样的场景和目标的刺激图像,被试在任务执行过程中需要大量感知和认知资源,行为表现较差且脑电信号的可分性较低。对过于简单的刺激图像,被试在任务执行过程中主要依靠感知功能,行为表现较好且脑电信号稳定。在实验设计中,适当控制场景内容或许是更合理的控制实验难度的手段。(3)构建一个针对车辆目标检测任务的刺激图像集并设计注意力引导策略,探索原生背景的场景复杂度和任务相关性对专家化视觉目标检测的影响。研究表明,自下而上和自上而下因素都会对视觉目标的检测产生影响。其中,适量的注意力线索和集中的任务相关注意力对自然场景中的目标检测是有益的,而过多的注意力线索和细粒度但语义无关的场景信息是无益的。该结果表明,对自然场景中视觉目标的高效处理可能涉及共存于自然场景中的语义和干扰信息之间的竞争过程,最终形成一个具有良好任务性能和高能效的视觉处理系统。本研究加深了我们对原生背景在目标检测中的作用的理解,为专家化视觉加工机制的研究提供了新的思路。更多还原

【Abstract】 Vision is an important way for humans and animals to perceive information from the outside world and is important for the survival and development of the organism.The study of human vision also has an important theoretical value for the development of computer vision.In computer vision research,understanding the working mechanism of human vision is an important step to achieve the goal of independent and automatic image processing,which is also one of the key and difficult points in this research area.However,most of the current computer vision approaches better simulate the low-level perceptual processes of human vision and do not provide a good representation of the equally important higher-level cognitive processes of human vision,which is mainly attributed to the insufficient understanding of the higher-level visual processing mechanisms of humans.Human processing of visual information is a complex process that is influenced by multiple bottomup stimulus drivers and top-down purpose-oriented factors,with significant individual specificity and a degree of commonality and similarity across visual domains.Research based on specific visual domains is an effective means to explore the domain-specific visual processing mechanisms,and can also provide an important theoretical basis for the study of the commonality of visual processing mechanisms among different visual domains.In the specific vision domain,the readers with expert visual ability have higher judgment accuracy,faster decision speed,more stable behavioral performance and stronger anti-interference ability compared with amateurs.Therefore,studies based on expert vision capabilities can yield more reliable and practically applicable results.The acquisition of expert visual abilities is associated with activity in a wide range of neural systems,involving not only the visual system but also extending to multiple higher functional areas throughout the cerebral cortex,but previous studies have typically described the effects of practice or experience on the functional representation of the brain during task performance in terms of changes in the increase,decrease,and functional reorganization of brain area activation,and have not elucidated the acquisition of expert visual abilities from a whole-brain perspective experiencing How plasticity in neural interaction patterns of the brain is modulated.Whole-brain analysis of neural interaction patterns in the brain requires a rational division of brain functional areas,and this division scheme is highly subjectspecific and domain-specific for expertized visual ability,which is one of the difficulties of this study.In addition,the functional linkage analysis of the whole brain is inevitably affected by spurious functional connectivity,and how to choose a reasonable functional connectivity analysis scheme is another difficulty of this study.Previous studies on top-down and bottom-up mechanisms in visual processing usually use synthetic or artificially adjusted visual stimuli,and while such visual stimuli can bring controlled variation and difficulty to the task,the neural coding about these factors is studied,but they cannot reflect or maintain the semantic effects in the real world and inevitably strip some unknown but important information from the native context.In this paper,we investigate expert visual mechanisms for real-world target detection based on vehicle detection tasks in native scenes.Real-world scenes are complex and diverse and contain a large amount of information,but the study of expertized vision mechanisms needs to be conducted based on stable and reliable brain responses,and the construction of its stimulus image set and the design of its experimental paradigm are highly task-relevant.Therefore,the experiments should retain useful information while avoiding the influence of distracting factors as much as possible.To address the above issues,this paper investigates the central neural mechanisms involved in the acquisition of expert visual ability,the construction of stimulus sets and experimental paradigm design for the study of expert visual mechanisms for target detection in native scenes,and the behavioral and neural performance of expert visual mechanisms for target detection under different task conditions,based on the imaging physician model and the vehicle detection model in native scenes,respectively.The main work and findings of this paper are summarized as follows.(1)Explored the mechanisms of modulation of functional connectivity patterns between brain regions in resting-state brain activity by radiographic visual experience using the model of radiologists and fMRI functional connectivity analysis.The findings suggest that radiographic visual experience may be accompanied by higher functional synchronization between brain regions responsible for perception,higher functional coupling between cognitive functions,and functional decoupling between perceptual and cognitive functions.This study fills a gap in the study of the dynamics of spontaneous connectivity between brain functional areas by radiographic visual experience,and may provide new ideas for understanding the central neural mechanisms underlying the formation of real-world visual recognition expertise.(2)Investigated the effects of image features on subjects’ behavioral and neural performance,and explored the design strategies for the construction scheme and experimental paradigm of stimulus image sets applicable to the study of this expert visual ability for vehicle target detection.Behavioral and EEG studies found that experimental designs for the study of neural mechanisms of visual target processing need to take into account the complexity of the stimuli.For complex and diverse stimulus images,subjects require a large amount of high-level cognitive resources during the task,and have poor behavioral performance and chaotic EEG signals.For overly simple stimulus images,subjects relied mainly on low-level cognitive functions during the task,with better behavioral performance and stable EEG signals.In addition,subjects’ familiarity with the stimuli had an important influence on individual behavioral performance during the task and the separability of the signals of brain responses,and subjects with expert visual experience were more suitable for the study of relevant visual processing mechanisms.(3)Explored the effects of scene complexity and task relevance of the native context on visual target detection using a new stimulus set and attention-guiding strategy.It is shown that both bottom-up and top-down factors have an impact on visual target detection.Among them,a moderate amount of attentional cues and focused task-relevant attention are beneficial for target detection in natural scenes,while too many attentional cues and finegrained but semantically irrelevant scene information are not beneficial.The results suggest that efficient visual processing of visual targets in natural scenes may involve a competitive process between semantic and distracting information coexisting in natural scenes,culminating in a visual processing system with good task performance and high energy efficiency.This study deepens our understanding of the role of native context in target detection and provides new ideas for the study of expert visual processing mechanisms.更多还原