Prediction of surface reflectance using a deep learning model trained on synthetic surface images

Abstract

Surface reflectance primarily depends on geometric properties. Although obtaining precise geometric details from real surfaces is challenging, deep learning models trained on synthetic surfaces with varying roughness can provide valuable insights. Such models, although they do not currently exist, might have the potential to accurately predict reflectance from real surface images that include roughness information. In this article, we propose a deep learning model based on Residual Network 34 to predict the reflectance of rough surfaces using confocal laser scanning microscopy (CLSM) surface images as the inputs. The model was trained using only synthetic surface images constructed by combining two basic patterns of inverted pyramids and cones. The ground-truth reflectance of the synthetic surfaces was obtained using ray tracing simulations combined with angle-dependent Fresnel reflection. The optimal synthetic surfaces were observed to have a pattern array size of 28 x 28 and an optimal pattern mixing ratio of 5:5. Thirty-two stainless steel surfaces with a wide reflectance range of 8.54-55.40% were fabricated by forming laser-induced periodic surface structures with a femtosecond laser. Three were used for validation, and the remaining 29 were used as the test dataset. When three actual surface CLSM images were used for validation, the R2 prediction accuracy of the model was 92.40%, and the mean absolute error of 29 tests was 2.98%. This study demonstrates that a model trained with synthetic surfaces can be used to predict the reflectance of real surfaces.