Lambertian Reflection Separation Under High Reflectiveness for Worn Surface Reconstruction With Insufficient Samples

Abstract

Compared with 2D images, 3D image based worn inspection can provide more profound information regarding the in-suit endoscope examination of mechanical parts. With photometric stereo vision, 3D topography can be reconstructed from 2D images with minimal cost. However, this promising technology is limited by the imprecion in calculating normal vector due to the high reflectiveness of worn surfaces. An innovative model is here developed to separate Lambertian reflection from photometric images. The Lambertian reflection variations are estimated by fusing multiple photometric images, while the constants are extracted from the general plane base of worn surfaces. Considering a small sample set of worn surfaces, hybrid samples, including the real and the simulated images, are adopted for parameter optimization with feature-level alignment. Moreover, a label-free prediction loss function is constructed to constrain the optimization direction via real samples. The constructed model is obtained from training 20,000 simulated samples and 1,000 real samples. The results indicate that high reflectiveness can be addressed by separating the Lambertian reflections. Comparatively, the model can improve the normal vector calculation with lower average error than other existing algorithms.