Neural network based photometric stereo for object with non-uniform reflectance factor

Proposes a new neural network-based photometric stereo approach for objects with a non-uniform reflectance factor using four input images acquired under different illumination conditions. The approach is empirical and uses the radial basis function (RBF) neural network to perform non-parametric functional approximation. We exploit the redundancy inherent in the four image irradiance values measured at each surface point in order to determine a local confidence estimate. This is achieved by training two distinct neural networks from a calibration sphere. The first neural network maps the image irradiance to both the surface normal and the reflectance factor (albedo). The second network maps the surface normal and the reflectance factor to the image irradiance. A comparison between the actual input and the inversely predicted input is used as the confidence estimate. Experiments on real data are described.