A signal-detection-based confidence-similarity model of face-matching

Face-matching consists of the ability to decide whether two face-images (or more) belong to the same person or to different identities. Face-matching is crucial for efficient face recognition, and plays an important role in applied setting such as passport control and eyewitness memory. However, despite extensive research, the mechanisms that govern face-matching performance are still not well understood. Moreover, to-date, many researchers hold on to the belief that match and mismatch responses are governed by two separate systems, an assumption that likely thwarted the development of a unified model of face-matching. The present study proposes a unified unequal variance confidence similarity signal-detection-based model of face-matching performance, one that facilitates the use of receiver operating characteristics (ROC) and confidence-accuracy plots analyses to better understand the relations between match and mismatch responses, and their relations to factors of confidence and similarity. The model can account for the presence of both within-identity and between-identity sources of variation in face recognition, and explains a myriad of face-matching phenomena, including the match-mismatch dissociation. The model is also capable of generating new predictions concerning the role of confidence and similarity and their intricate relations with accuracy. The new model was tested against six alternative competing models (some postulate discrete rather than continuous representations) in three experiments. Data analyses consisted of hierarchically-nested model fitting, ROC curve analyses, and confidence-accuracy plots analyses. All of these provided substantial support in the signal-detection-based confidence-similarity model. The model suggests that the accuracy of face-matching performance can be predicted by the degree of similarity/dissimilarity of the depicted faces and the level of confidence in the decision. Moreover, according to the model confidence and similarity ratings are strongly correlated.人脸匹配（Face-matching）指的是判定两张（或多张）人脸图像是否属于同一身份或不同个体的能力。人脸匹配对于高效人脸识别至关重要，并在护照查验、目击者记忆等实际应用场景中发挥着关键作用。 然而，尽管已有大量相关研究，调控人脸匹配表现的内在机制仍未得到充分阐释。此外，迄今为止，诸多研究者仍秉持一种观点：匹配与不匹配反应由两套独立的认知系统支配，这一假设大概率阻碍了人脸匹配统一模型的研发。 本研究提出了一种基于置信度与相似性的不等方差信号检测统一模型，该模型支持运用受试者工作特征（ROC）曲线与置信度-准确率图开展分析，以更深入地理解匹配与不匹配反应之间的关联，以及二者与置信度、相似性因素的关系。 该模型能够解释人脸识别中同一身份内与不同身份间的变异来源，并可阐释包括匹配-不匹配分离效应在内的多种人脸匹配现象。 此外，该模型还可生成关于置信度与相似性的作用，以及它们与准确率之间复杂关联的全新预测。 本研究通过三项实验，将该新模型与六种竞争性替代模型（部分模型假设采用离散表征而非连续表征）进行了对比验证。 数据分析涵盖分层嵌套模型拟合、ROC曲线分析以及置信度-准确率图分析三类方法。 所有分析均为基于信号检测的置信度-相似性模型提供了显著的支持证据。 该模型表明，人脸匹配表现的准确率可通过所涉人脸的相似/不相似程度，以及决策时的置信度水平进行预测。此外，根据该模型，置信度评分与相似性评分之间存在强相关性。