Hide and Seek: Using Masked Vision Transformer to Detect Surface Structures on Laser Polished Metals

Polishing metallic materials with laser radiation (LP) is based on the melting of a thin surface layer. In the molten phase, surface roughness is smoothed because of the interfacial tension and the material solidifies with a smoothed surface. It also creates surface structures, which introduce roughness in a specific wavelength range. In most cases, multiple surface structures occur simultaneously. Altogether, they present a significant obstacle in achieving polishing results necessary for use in industry (Metric Ra often smaller 100 nm). However, once successfully detected, the role of surface structures towards roughness may be systematically reduced [1,2]. During process parameter development, surface struc tures must be analyzed manually by highly skilled process engineers using various, expensive analysis techniques to optimize process parameters (E.g., white-light interferometry and scanning electron mi croscope). To automate this process, this work evaluates state-of-the-art image processing methods to detect surface structures solely based on white-light interferometry. Therefore, process experts have identified up to eight different surface structures in over 2,500 white-light interferometry images to build a meaningful and representative dataset for surface structure classification on various materials. We benchmarked state-of-the-art machine learning models and show that both, ResNet [3] and vision transformer (ViT) [4] models are suitable techniques for classifying surface structures, achieving up to 82% accuracy on our dataset. Furthermore, we explore the self-supervised learning approach data2vec [5] to make unlabeled data usable by pretraining in a self-supervised fashion and show that features learnt are already descriptive enough to distinguish surface characteristics of different types without requiring any annotation. Building on state-of-art techniques, we propose a novel masking strategy to further improve the quality of the learnt features regarding surface properties, which may be beneficial in a much broader context than just metallic surfaces. With this new technique, we build a bridge for vast, unlabeled data which is often collected in large quantities from industrial machines but cannot be used in supervised machine learning without prior manual labelling.