Deep Network Based Estimation of Perceptual Surface Quality

Abstract

How we perceive quality of surface with various geometry and reflectance under various illuminations is not fully understood. One of widely studied approaches in understanding human perception is to derive image statistics and build up a model that estimates human perception of surface quality attributes. This work presents estimation of surface quality based on machine learning. Instead of deriving image statistics and building up estimation models, we use deep networks that can estimate perceptual surface quality directly from a surface image. The networks are trained from perceptual lightness and glossiness data obtained from psychophysical experiments. The performance of the networks is compared to image statistics derived from regression analysis. The trained deep networks provide estimation of surface quality with good correlation to human perception.