Psychophysical Studies of Recognition for Psychology and Computer Vision

Recognition is a core process in both human cognition and Computer Vision. Psychophysical studies provide a powerful framework for understanding visual recognition by precisely measuring behavioral responses to controlled stimuli. These experiments deepen our understanding of perception and categorization in psychology, while also offering valuable benchmarks for evaluating computational models in Computer Vision and Machine Learning.

This dissertation covers topics ranging from the use of human perceptual data in machine learning contexts to studying perception for theoretical insight, including comparisons with models to assess how well they capture human recognition behavior. We begin by exploring how human perceptual responses can guide improvements in classification models. We then analyze human eye gaze during face recognition tasks to examine attention and feature importance. Finally, we study how people perceive category boundaries under different stimulus distributions, comparing their behavior to models based on Extreme Value Theory (EVT) and Gaussian assumptions. Together, these studies emphasize the value of psychophysical insights for understanding recognition in both humans and Computer Vision models.

识别是人类认知与计算机视觉（Computer Vision）中的核心过程。心理物理学研究通过精准测量受控刺激下的行为反应，为理解视觉识别提供了强有力的研究框架。这类实验既深化了心理学领域对知觉与分类的认知，也为计算机视觉与机器学习（Machine Learning）领域的计算模型评估提供了极具价值的基准测试数据集。
本论文的研究主题涵盖从机器学习场景下人类知觉数据的应用，到以理论视角开展知觉研究，其中包含通过与模型对比以评估其对人类识别行为拟合程度的相关内容。本研究首先探讨人类知觉反应如何指导分类模型的优化；随后分析人脸识别任务中的人类眼动轨迹，以考察视觉注意力与特征重要性；最后研究不同刺激分布下人类对类别边界的知觉表现，并将其行为与基于极值理论（Extreme Value Theory，EVT）和高斯假设构建的模型进行对比。综上，这些研究均凸显了心理物理学视角在理解人类与计算机视觉模型识别机制中的重要价值。