Neural Network Based Photometric Strereo Using Illumination Planning

Abstract

This paper proposes a new approach of neural network based photom etric stereo in which all the directions of illumination are close to and is rotationally moved around the viewing direction . Adjace nt light sources cause a num erically ill-conditioned problem. Ill-conditioning is overcome in two ways. First , many more than thr ee images are acquired. Second , principal components analysis is used as a linear preprocessing technique to determine a r educed dimensionality subspace for use as input. Th e approach is empiri cal and us es the radial basis function (RBF) neural network to do non-parametric functional approximation. One neural network maps image irradiance to surface normal. A second network maps surface normal to image irradiance. Th e two networks are trained using samples from a ca libr ation sphere. Comparison between th e actual input and the inversely predicted input is used as a confiden ce est imat e. Further , we introduc e th e illumination planning , which dispos es the information of th e cast shadow region from th e ill-effected lig ht sources for each point on th e object using the confidence estimate. Th en , the best combination of NN units with well-effected light sources for each point are used to get the relevant results . The results on real data are also eva luat ed.