The tuning of face pareidolia to orientation statistics

Face pareidolia is a ubiquitous phenomenon in which observers detect face-like structures in natural textures, arrangements of common objects, or other patterns. Pareidolic face detection likely reflects observers’ internal representations of typical face structure and so may be tuned to the same low-level image features as typical face detection and recognition. Presently, we investigated whether face pareidolia, like other face recognition tasks, is tuned to horizontal orientation energy. We presented participants (N=43) with a series of fractal noise images (8 unique images of 1/f noise per condition) and asked them to report any pareidolic faces they saw in these patterns. To vary orientation statistics, we first generated isotropic noise images that were not biased in favor of any orientation passband and then applied orientation filtering to create horizontal and vertical versions of each stimulus. These images were printed on paper and presented to participants in a counterbalanced blocked design. Participants were instructed to examine each image and indicate the presence of any pareidolic faces by circling them using a marker, including a cartoon hat to indicate the orientation of the pareidolic face on the page. We predicted that observers would find more pareidolic faces in isotropic and horizontally-filtered images than vertically-filtered images. We counted the pareidolic faces detected in each image by our participants and analyzed this count data using a mixed-model analysis that included orientation as a fixed effect and participant as a random effect. This analysis revealed significantly negative slopes associated with the horizontal and vertical conditions relative to the isotropic condition, indicating that orientation filtering reduces the rate of pareidolic face detection, but vertical filtering is no more detrimental than horizontal filtering. Pareidolic face detection may thus depend on richer image structure (broadband orientation energy in particular) than mechanisms for face identification and categorization.