Robust visual recognition of high-level form in human observers

视觉识别是困难的。物体在不同的光照、观看距离和方向条件下被看到，这给它们的图像带来了巨大的变化。此外，观察者因素——从眼睛和大脑引入的扭曲到观察者寿命期间不同的视觉经验——都会影响视觉识别。尽管取得了重大进展，但对这些因素在视觉识别计算中的影响的全面理解仍然难以捉摸。我们最近的工作表明，视觉系统在视觉输入中利用持续的规律，例如，尺寸或表情，来促进对面孔的识别。在一项自然观察研究中，我们通过嵌入眼镜的摄像头获取了日常视觉体验的镜头。对这段视频的分析显示，大多数人看到的人脸都是来自社交距离，也就是说，视觉上比较大。事实上，我们发现在社会交往中最常见的脸的大小与更好的识别有关。我们对大小和模糊规律的研究，也发现了与该提议的总体假设一致的证据，即视觉系统采用基于环境中最常见的视觉输入类型的策略来优化识别性能。我研究的长期目标是：(1)描述人类观察者视觉输入中的各种统计规律及其对识别的影响，包括物理规律，例如来自光学几何的规律，以及观察者与环境相互作用方式的结果；(2)揭示利用这些规律促进识别的视觉系统策略；(3)开发新的应用程序，如计算机程序；为人类视觉系统量身定制的，以增强在低可见度条件下的识别，例如对颗粒状图像的识别。为了追求上述长期目标，我的短期目标侧重于以下方面产生的特定统计规律：(1)与面部的观看距离（例如，视觉尺寸和模糊度），(2)社会环境中最常见的面部的种族，以及(3)整体面部暴露时间。在接下来的5年周期中，我们将继续依靠我们在方法方面的专业知识，例如行为测试（例如，视觉心理物理学），计算建模（例如，理想观察者）和视觉图像统计的自然观察（例如，基于通过嵌入眼镜的摄像机获取的镜头）。各级高素质人才将获得科学和技术技能，这些技能将为他们未来在学术界、工业界和医学界的职业生涯服务。这项工作将促进我们对大脑视觉识别基本原理的理解，并为开发工具提供信息，以提高在各种环境下（从计算机应用到低能见度条件下的视觉辅助）对高级形式的识别。

Visual recognition is hard. Objects are seen in a variety of conditions differing in illumination, viewing distance, and orientation, which introduce drastic changes to their image. In addition, observer factors-from distortions introduced by the eye and the brain to varying visual experience over the observer's lifespan-influence visual recognition. Despite significant advances, a comprehensive understanding of the impact of these factors in computations underlying visual recognition remains elusive. Our recent work showed that the visual system takes advantage of sustained regularities, e.g., those in size or expression, in the visual input to facilitate recognition of faces. In a naturalistic observation study, we acquired footage of daily visual experiences via eyewear-embedded cameras. Analysis of this footage revealed that most views of faces are from social interaction distances, i.e., visually large. Indeed, we found that face sizes most common in social interactions are associated with better recognition. Our work on regularities in size and blur, also uncovered evidence consistent with the overarching hypothesis of this proposal that the visual system adopts strategies to optimize recognition performance based on the most frequent types of visual input in the environment. Long-term objectives of my research are: (1) Describe various statistical regularities in human observers' visual input and their impact on recognition, including physical regularities, such as those that arise from the geometry of optics, as well as those that are consequences of the ways in which observers interact with their environment, (2) Uncover strategies of the visual system that utilize these regularities to facilitate recognition, and (3) Develop novel applications, such as computer programs, tailored to the human visual system to enhance recognition in low visibility conditions, e.g., of grainy images. My short-term objectives focus on specific statistical regularities that arise from (1) viewing distance to faces (e.g., visual size and blur), (2) ethnicities of faces most prevalent in the social environment, and (3) overall face exposure duration, in pursuit of the long-term objectives above. In the next 5-year cycle, we will continue to rely on our expertise in methodologies such as behavioural testing (e.g., visual psychophysics), computational modelling (e.g., ideal observers), and naturalistic observation of visual image statistics, (e.g., based on footage acquired through eyewear-embedded cameras). Highly qualified personnel at all levels will gain scientific and technical skills that will serve them in future careers in academia, industry and medicine. This work will advance our understanding of the fundamental principles underlying visual recognition in the brain and inform the development of tools that improve recognition of high-level form in a variety of settings from computer applications to visual aids in low-visibility conditions.