Decoding the public gaze: a deep learning framework for assessing the perception of architectural heritage

Approaches to urban heritage governance increasingly prioritise material form and quantifiable metrics, yet how measurable environmental quality relates to the public’s evaluative-affective experience of heritage remains poorly understood. This study develops a perception-sensitive diagnostic framework that integrates deep learning on street-view imagery with Multiscale Geographically Weighted Regression to examine the spatial congruence and divergence between objective environmental indicators and computationally inferred perceptual responses. Applied to the Xi’an City Wall, the analysis reveals pronounced spatial heterogeneity and identifies two diagnostic zone types: resonance zones, where positive perceptual responses persist despite modest environmental quality, and discordance zones, where high environmental quality fails to elicit correspondingly positive scores. MGWR results show that the influence of factors such as road density and green-space coverage varies substantially in direction and magnitude across the study area, confirming that globally uniform assessment assumptions obscure locally important dynamics. By rendering these patterns visible, the framework contributes a scalable diagnostic layer to the perceptual mapping called for by the Historic Urban Landscape approach and provides empirical grounding for the argument that technocratic assessment may systematically differ from resident lived experience, supporting more perception-sensitive and spatially targeted heritage management.