Surface roughness classification using light scattering matrix and deep learning

High-quality optical components have been widely used in various applications; thus, extremely high beam quality is required. Moreover, surface roughness is a key indicator of the surface quality. In this study, the angular distribution of light scattering field intensity was obtained for surfaces having different roughness profiles based on the finite difference time domain (FDTD) method, and the results were compared with those obtained using the generalized Harvey-Shack (GHS) theory. It was shown that the FDTD approach can be used for an accurate simulation of the scattered field of a rough surface, and the superposition of results obtained from many surfaces that have the same roughness level was in good agreement with the result given by the analytic GHS model. A light scattering matrix (LSM) method was proposed based on the FDTD simulation results that could obtain rich surface roughness information. The classification effect of LSM was compared with that of the single-incidence scattering distribution (SISD) based on a ResNet-50 deep learning network. The classification accuracy of the model trained with the LSM dataset was obtained as 95.74%, which was 23.40% higher than that trained using the SISD dataset. Moreover, the effects of different noise types and filtering methods on the classification performance were analyzed, and the LSM was also shown to improve the robustness and generalizability of the trained surface roughness classifier. Overall, the proposed LSM method has important implications for improving the data acquisition scheme of current light scattering measurement systems, and it also has the potential to be used for detection and characterization of surface defects of optical components.