Supervised learning for roughness reconstruction under different scanning modes using a confocal laser scanning microscope

The application of artificial intelligence (AI) to address the trade-off between acquisition speed and precision in industrial confocal laser scanning microscopy (CLSM) remains largely unexplored. This study presents a systematic framework to tackle this challenge. A dedicated network, termed "leaky ReLU residual-in-residual neural network" (LRIRN), is developed to enhance the signal-to-noise ratio (SNR), supported by a dataset acquired using two scanning modes: fast scanning mode (FaSM) for low-SNR data and fine scanning mode (FiSM) for high-SNR reference data. A task-specific loss function is designed to guide the optimization process, and a comprehensive evaluation scheme is established, incorporating two novel, to the best of our knowledge, metrics alongside surface roughness and image restoration indicators. Visualization maps and cross-sectional profiles, together with quantitative assessments, are employed for both qualitative and quantitative validation. Experimental results reveal that the proposed framework achieves an SNR in FaSM-acquired data comparable to FiSM, while effectively compensating for displacement deviations in the <i>y</i> and <i>z</i> directions. Our findings highlight the potential of the proposed AI-driven method to enhance the efficiency and accuracy of CLSM measurements, paving the way for high-speed, high-precision industrial surface metrology.