A novel surface normal estimation method using smooth L1 regression for photometric stereo

Photometric stereo reconstructs surface shape with intricate details by analyzing the light propagation process and has attracted wide range of applications such as industrial measurement. However, the presence of non-Lambertian reflections on real-world scenarios poses a significant challenge to surface normal estimation methods. Most of the existing approaches aim to achieve reasonably accurate results by filtering non-Lambertian observations through employing iterative frameworks. Based on these, this study introduces a novel surface normal estimation method that formulates the observation selection and regularization as a smooth <i>L₁</i> regression problem. Specifically, we sort pixels by intensities and select effective observations through a threshold strategy, the surface normal are then estimated by a smooth <i>L₁</i> loss function to resist non-Lambertian corrosions so that facilitate a more accurate result. The performance of the method is validated through testing on real-world datasets, with an impressive average angular error as low as 11.92&deg;. In experiments, surface reconstruction of a turbine blade is successfully achieved, showcasing its applicability in industrial manufacturing.