Illumination Planning for Measuring Per-Pixel Surface Roughness

Measuring per-pixel surface roughness is useful for machine vision applications such as visual inspection. The surface roughness can be recovered from specular reflection components, but a large number of images taken under different lighting and/or viewing directions is required in general so that sufficient specular reflection components are observed at each pixel. In this paper, we propose a robust and efficient method for per-pixel estimation of surface roughness. Specifically, we propose an illumination planning based on noise propagation analysis; it achieves the surface roughness estimation from a small number of images taken under the optimal set of light sources. Through the experiments using both synthetic and real images, we experimentally show the effectiveness of our proposed method and our setup with a programmable illumination and a polarization camera.