High-speed and High-accuracy Nanoprofile Reconstruction by Fusing Lateral Scanning Optical Interferometry and Deep Learning-driven Image Registration

Abstract Lateral scanning white-light interferometry is a promising noncontact technique that offers a wide field of view (FOV) for high-precision surface topography measurement. However, during scanning, nonuniform motion in the linear translation stage induces lateral displacement error, resulting in pattern misalignment and distortion of the reconstructed surface topography. This study presents a recovery algorithm based on deep learning-driven image registration. The SuperPoint and SuperGlue techniques were used to detect feature points in adjacent interferograms and estimate the actual lateral displacements. The lateral displacement error was calculated as the accumulated difference between the measured lateral displacement and the theoretical sampling interval, and it was sequentially compensated to produce corrected interferograms. The phase-shift error was estimated by extracting the amplitude envelope and identifying the envelope peak based on planar regions of each interferogram. Spatial sampling points were corrected using a spline interpolation method to obtain uniformly sampled interference signals. Finally, the surface topography was reconstructed by integrating a moving sine wave fit with a seven-step phase-shifting method. The experimental results demonstrate that the proposed method effectively corrects lateral displacement errors of periodic nanoscale surface structures during high-speed scanning.