Deep-learning based fringe-print-through error and noise removal for dynamic phase-shifting interferometry

High-speed and precise surface measurement is crucial in manufacturing, particularly in the semiconductor industry. Dynamic phase-shifting interferometry is a highly efficient and widely recognized optical metrology technique, known for its exceptional accuracy and speed, making it ideal for industrial inspection and measurement tasks. However, incorrect phase-shift intervals between interferograms generated by this technique can lead to fringe-print-through (FPT) errors in the surface measurements. Additionally, additive Gaussian noise present in the interferograms complicates the accurate assessment of residual surface after measurements. Rapidly eliminating these FPT errors and noise is essential for achieving high-accuracy and high-speed measurement applications. In this paper, we propose a novel deep-learning method to simultaneously eliminate FPT errors and noise in dynamic phase-shifting interferometry. Our approach utilizes a UNet++ deep-learning network, which processes the surface phase containing errors as input and outputs the corresponding FPT errors and noise. Trained on simulated data, the model learns to directly predict these errors and noise from the surface phase with errors. Consequently, the corrected surface phase is obtained by subtracting the predicted FPT errors and noise from the initial surface phase. Simulation and experimental results demonstrate that our deep-learning method effectively removes FPT errors and noise, providing broad versatility, rapid processing, and robustness, thereby significantly enhancing measurement accuracy in dynamic measurement applications.