Scene Structure Predicts Perceptual Decisions in Naturalistic Detection Tasks

Abstract

The human visual system can identify objects in complex natural scenes, yet the mechanisms supporting robust perception under such variable conditions remain incompletely understood. Here, we investigate how the statistical structure of natural scenes shapes perceptual evidence formation and determines whether near-threshold stimuli are perceived correctly or incorrectly. We combine controlled psychophysics, outdoor augmented reality (AR), deep neural networks (DNNs), image-feature analysis, and EEG to examine how background context modulates perceptual decisions. Across multiple detection tasks, human performance was systematically influenced by probe-free background structure. DNNs trained on background images alone predicted correct and incorrect behavioral outcomes, with stronger effects in postcue conditions, suggesting that global scene context contributes to local perceptual decisions when spatial uncertainty is higher. AR experiments further showed that these context-driven effects persist in naturalistic viewing environments. To identify the visual information underlying these effects, we analyzed low-level image statistics. Texture entropy and edge density emerged as informative features, and conventional machine-learning models trained on these measures achieved meaningful correct/incorrect classification. Finally, EEG analyses revealed neural signatures of image-driven perceptual variability: activity during the probe-free preimage window distinguished later correct from incorrect trials, and combining EEG with image-derived features improved decoding performance. Together, these findings show that perception in natural scenes is not determined solely by the target, but is shaped by the statistical structure of the surrounding context. By linking psychophysics, AR, DNN modeling, image statistics, and EEG, this work provides a unified framework for understanding how environmental structure and neural dynamics jointly support perceptual decision-making.