Non-Lambertian photometric stereo by color and noise analysis

This paper proposes a novel photometric stereo method for reliable reconstruction of non-Lambertian object surfaces based on color and noise analysis. Unlike previous techniques that treat highlight pixels as outliers, the proposed method utilizes the intrinsic shape cue in those pixels by robust reflection separation. Image noise is estimated from two photometric images instead of a single one, thus the common uniform block assumption can be eliminated. The highlight pixels are identified by analyzing intensity and chromaticity uncertainty, and the specularity components are removed based on robust estimation of diffuse chromaticities. The normals of deficient positions, whose pixels are of extremely low intensities, are computed from individual channels. The obtained surface normal map is further refined by adaptive lateral filtering. Experimental results show that, compared with the state-of-the-art techniques, the proposed method produces improved reconstruction accuracy, which indicates the effectiveness of the proposed method for non-Lambertian object surfaces.