BlurRes-UNet: A novel neural network for automated surface characterisation in metrology

Abstract

Surface characterisation is essential in metrology for precise measurement and analysis of surface features, ensuring product quality and compliance with industry standards. Form removal is the primary step in surface characterisation, isolating features of interest by eliminating the primary shape from measurements. Traditional least-squares methods, as specified in ISO standards, are effective but offer limited adaptability for diverse surfaces and often require manual parameter tuning. With this limitation in mind, this paper proposes BlurRes-UNet, a deep learning-based model designed for fully automatic form removal. Built on an encoder–decoder architecture with residual learning, skip connections, and a tailored loss function, the model incorporates domain knowledge, feature engineering, and regularisation techniques to optimise performance with limited training data. The model is evaluated against traditional least squares methods and assessed using various strategies to demonstrate its performance and robustness. It processes surfaces of 256 × 256 resolution in 7.32 ms per sample on a T4 GPU, achieving superior accuracy in recognising reference forms across diverse surfaces compared to traditional methods. The results suggest that the model is capable of accurately recognising different order reference forms from diverse surfaces, facilitating an autonomous surface characterisation system without the need for manual intervention.