Variability as a route to understanding face recognition

这个项目代表了一种看待人脸识别问题的新方法。尽管对这一主题进行了大量的研究，但我们仍然不了解人脸处理的最基本方面：我们如何识别我们看到的人？这是人类感知中的一个关键问题，但它在法医和安全环境中也具有实际意义。这个项目源于一个简单的观察：同一张脸的照片可能看起来非常不同。在人脸识别的标准方法中，这一司空见惯的事实被视为一种不便。同一个人的照片之间的差异被视为“噪音”，通过系统地控制用于研究的图像来忽略或消除。这个研究项目采取了完全相反的方法。而不是试图控制这种可变性，它将被明确地研究。在这种方法下，人脸识别的问题不是如何“区分人”，而是如何“区分人”——如何将表面上不同的图像组合成一个连贯的表示。早期的研究表明，熟悉面孔识别的一个非常重要的组成部分是概括表面图像差异的能力——这些差异往往会欺骗不熟悉的观众，以及基于计算机的自动系统。目前未能解决这种可变性可能是人脸识别进展缓慢的原因——这一进展已经落后于对人脸处理的其他方面（如社会感知）的理解。本研究由三个部分组成。首先，系统地检查同一个人的图像之间的物理差异。将统计技术应用于图形数据，目的是指定面部图像的哪些方面通常会发生变化，以及哪些方面对该个人有特殊变化。其次，一系列的行为实验将检验我们对熟悉面孔的表征的本质——假设是这种表征需要包含可变性。第三，一系列研究将解决人类观察者的实际人脸识别问题（例如，出于安全目的）。基于计算机的方法也将在这方面进行审查，因为这些系统仍然非常差，尽管供应商声称。这种新颖的人脸识别方法有可能为近年来进展相当缓慢的领域做出重大贡献。

This project represents a new way to look at the problem of human face recognition. Despite a large amount of research on this topic, we still do not understand the most fundamental aspect of face processing: how can we identify the people we see? This is a key problem in human perception, but it also has practical implications in forensic and security settings. This project has its roots in a simple observation: pictures of the same face can look very different indeed. In the standard approach to face recognition, this commonplace fact is treated as an inconvenience. Differences between pictures of the same person are regarded as 'noise', and either ignored or eliminated by systematically controlling the images used for research. This research programme takes exactly the converse approach. Instead of trying to control away this variability, it will be studied explicitly. Under this approach, the problem of face recognition is not how to 'tell people apart', but instead how to 'tell people together' - how to bring together superficially different images into a coherent representation. Early work suggests that a very important component of familiar face recognition is the ability to generalize over superficial image differences - differences which tend to fool unfamiliar viewers, as well as automatic computer-based systems. The current failure to address this variability may account for the slow progress in face identification - progress which has fallen behind the understanding of other aspects of face processing such as social perception. This research comprises three components. First, a systematic examination will be conducted of the physical differences between images of the same person. Applying statistical techniques to graphical data, the aim is to specify what aspects of face images vary commonly, and what aspects vary idiosyncratically to that person. Second, a series of behavioural experiments will examine the nature of our representations of familiar faces - the hypothesis is that this representation needs to incorporate variability. Third, a series of studies will address practical face recognition by human observers (e.g. for security purposes). Computer-based approaches will also be examined in this strand - as these systems remain very poor, despite the claims of vendors. This novel approach to face identification has the potential to make a significant contribution to an area which has progressed rather slowly in recent years.