Friday, November 20, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News

Deep Learning Registration Corrects Scanning Errors in High-Speed Optical Nanoprofiling

November 20, 2026
in Science News
Blake Davidson
By Blake Davidson Scienmag Editorial Profile - Data Science
Reading Time: 6 mins read
0
Deep Learning Registration Corrects Scanning Errors in High-Speed Optical Nanoprofiling

Deep Learning Registration Corrects Scanning Errors in High-Speed Optical Nanoprofiling

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

Optical metrology has emerged as a cornerstone of precision manufacturing, enabling noncontact three-dimensional characterization of microscale and nanoscale structures across photonics, microfluidics, and biomedical instrumentation. Extending this field of view historically required multi-tile image stitching routines, which routinely introduce geometric distortions and severe pattern misalignment along stitched seams. To eliminate the inefficiencies of vertical translation, lateral scanning white-light interferometry tilts the optical interferometer relative to the surface normal, creating an inclined coherence plane that allows high-throughput surface mapping through a single lateral movement. However, high-speed lateral scanning introduces physical obstacles: motorized linear translation stages suffer from mechanical straightness deviations and velocity instability. These physical fluctuations introduce lateral displacement errors between successive imaging frames, triggering spatial pattern distortion and fringe shifts that derail surface topography reconstruction algorithms.

To overcome these mechanical bottlenecks without requiring complex and bandwidth-limited hardware compensation stages, researchers have developed an integrated computational framework that couples lateral scanning white-light interferometry with deep learning-driven image registration. The technique corrects lateral displacement errors and phase-shift deviations caused by translation stage imperfections directly from acquired interferograms. By deploying neural feature extraction and attention-based correspondence matching across consecutive camera frames, the system reconstructs nanoscale profiles with high fidelity during continuous motion. In physical validation trials, the system achieved a lateral scanning velocity of approximately one millimeter per second while maintaining vertical measurement accuracy comparable to standard commercial instruments. The reconstruction workflow integrates deep learning registration, one-dimensional linear resampling, planar envelope extraction using moving sine wave fitting, spline interpolation, and a seven-step phase-shifting algorithm to isolate surface elevations with sub-nanometer consistency across periodic nanostructured specimens.

The physical imaging apparatus centers on a custom lateral scanning white-light interferometer illuminated by a blue light-emitting diode featuring a central wavelength of 455 nanometers and a spectral bandwidth of 18 nanometers. The beam passes through a Köhler illumination assembly before a beam splitter divides it into reference and object paths. Both arms incorporate identical 50-power microscope objectives with a numerical aperture of 0.50, focusing illumination onto an adjustable reference mirror and the translating target specimen. The target rests on a motorized linear translation stage programmed to scan horizontally beneath the tilted optical column. In the ideal theoretical model, the spatial sampling interval between consecutive acquired frames depends strictly on stage velocity, optical magnification, pixel size, and camera frame rate. Any momentary speed jitter instantly violates this integer sampling relationship, scattering sub-pixel positioning errors throughout raw interferogram stacks.

Investigating the primary physical root of data distortion during high-speed scans revealed that the camera maintained an exceptionally steady acquisition cadence of 1000 frames per second without frame loss or temporal dropouts. In contrast, the linear motorized translation stage exhibited velocity variations exceeding one percent around its target speed of 0.998 millimeters per second. This confirmed that motorized speed instability dominates spatial sampling errors during continuous motion. Traditional handcrafted feature detection methods, including scale-invariant feature transform, oriented FAST and rotated BRIEF, and accelerated-KAZE, frequently break down when analyzing raw interferograms due to repetitive optical fringe patterns, smooth backgrounds, and low regional texture contrast. To establish robust frame-to-frame correspondence despite fringe interference, the researchers implemented a combined deep learning architecture using SuperPoint and SuperGlue to extract descriptors and identify true matching coordinates across neighboring frames.

The feature detection pipeline relies on the SuperPoint convolutional network, which was trained on synthetic geometric shapes featuring well-defined corners, such as triangles, quadrilaterals, line segments, and cuboids, embedded with synthetic noise and optical blur. Because these synthetic images were generated algorithmically, exact ground-truth vertex coordinates were known without human annotation. The resulting base detector subsequently processed real images through homographic adaptation, applying random homographic transformations across scale, rotation, and viewing angle to produce robust pseudo-labels for self-supervised weight refinement. When applied to real lateral interferograms, the refined SuperPoint model extracts interest points and high-dimensional local descriptors. These representations are fed into SuperGlue, a graph neural network that leverages self-attention and cross-attention mechanisms to combine keypoint locations, visual descriptors, and match confidence scores. This deep learning approach produces dense, highly reliable keypoint correspondences across adjacent interferograms where classical detectors fail.

By comparing the predicted pixel correspondences from SuperGlue against the nominal optical sampling interval, the algorithm determines the actual physical displacement occurring between sequential camera exposures. Because stage velocity fluctuates dynamically, sampling error accumulates over time relative to the initial frame of the sequence. The algorithm defines the total lateral displacement error for any given frame as the cumulative difference between actual measured lateral shifts and the theoretical integer sampling interval. Once calculated, this accumulated displacement error is sequentially compensated across the discrete horizontal coordinate grid. Because downstream signal processing algorithms require uniformly distributed integer-indexed coordinates, a one-dimensional linear interpolation resamples each interferogram along the corrected spatial coordinates. Because continuous linear resampling progressively contracts the effective field of view along the lateral scanning axis, the final frames of the recorded series are cropped from the rearranged dataset to ensure uniform spatial bounds.

Compensating for horizontal translation errors does not entirely resolve optical distortion, because physical straightness deviations in the motorized translation stage alter the vertical clearance between the optical head and the specimen. This vertical deviation produces lateral fringe shifts between successive frames, altering the zero optical path difference position and corrupting the phase information required for exact depth calculation. To isolate phase-shift error without interference from structural topography, the algorithm detects planar reference zones on the sample surface. Structural surface patterns introduce intrinsic phase shifts during lateral movement, meaning patterned zones must be separated from true stage-induced fringe movements. The algorithm applies a moving sine wave fitting algorithm to extract the amplitude envelope from local interference signals, fitting the resulting envelope with a parabolic function. By evaluating the root-mean-square error between the empirical envelope and the fitted parabola, the algorithm segments flat planar surfaces from patterned zones through threshold testing.

Once planar zones are isolated, the system calculates the frame-to-frame phase-shift error relative to the initial frame using a conventional planar white-light phase-shifting interferometry routine. Spectral analysis of these phase-shift variations revealed low-frequency mechanical fluctuations concentrated entirely within the zero to 50 hertz bandwidth, reflecting physical stage guidance imperfections. After correcting lateral displacement errors, the recorded interferograms are mathematically rearranged into a sequence corresponding to individual sample points moving through the coherence plane. Applying the measured phase-shift compensation produces discrete, non-uniformly spaced temporal interference signals for each physical coordinate. Because phase-shifting reconstruction equations assume uniform sampling intervals, the discrete signals are reordered and resampled using one-dimensional spline interpolation, yielding smooth, uniformly distributed interference wavepackets that exhibit markedly superior regularity compared to raw uncorrected signals.

Final three-dimensional topography reconstruction utilizes a hybrid signal processing framework that merges moving sine wave fitting envelope extraction with a seven-step phase-shifting algorithm. The amplitude envelope of each corrected interference signal is calculated from in-phase, quadrature, and background intensity components. Fitting a parabolic curve across the localized envelope isolates its vertex, pinpointing the peak position of zero optical path difference. The absolute surface height for each pixel is subsequently determined through an analytical expression combining the central light-emitting diode wavelength, the envelope peak coordinate, and the evaluated phase angle, effectively eliminating residual zero optical path difference errors.

Scanning across a total acquisition time of 0.5 seconds at 1000 frames per second yielded 500 raw interferograms, forming a rearranged field of view 250 micrometers in length. Without deep learning registration, the reconstructed surface exhibited severe geometric distortion, elliptical pattern deformation, and jagged boundary artifacts caused by translation stage jitter.

Quantitative accuracy was benchmarked by comparing profile measurements of the etched array across five sample groups against a commercial vertical scanning white-light interferometer serving as an established reference standard. Across the five test regions, the proposed lateral scanning system measured an average etched step height of 210.525 nanometers with a standard deviation of 0.212 nanometers, while the commercial instrument recorded an average depth of 210.483 nanometers with a standard deviation of 0.130 nanometers. The absolute measurement discrepancy between the two techniques remained below 0.05 percent, with an overall standard deviation of less than 0.5 nanometers. Computational processing demanded 235.65 microseconds per single-point interference signal, compared to 57.63 microseconds for standard Hilbert transforms and 79.32 microseconds for fast Fourier transform approaches on uncorrected signals, representing a modest computational trade-off for substantial accuracy gains during high-speed translation.

To evaluate performance on non-planar curved geometries, the system profiled a commercial cylindrical lens array exhibiting a nominal pitch of 400 micrometers and a nominal radius of curvature of 0.7 millimeters. Reconstructing the sectional curvature profile along the scanning axis produced a measured average radius of curvature of 0.708 millimeters, deviating by less than 0.01 millimeters from the nominal specification. Future development will focus on establishing generalized algorithmic frameworks capable of high-speed, in-line industrial metrology across complex, non-periodic micro-surfaces without requiring planar reference constraints.

Subject of Research: Lateral scanning optical interferometry and deep learning image registration for nanoscale surface metrology

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

Article References: Wu, Q., Zhao, H., Li, Z., Wang, J., & Zhu, J. (2025). High-speed and High-accuracy Nanoprofile Reconstruction by Fusing Lateral Scanning Optical Interferometry and Deep Learning-driven Image Registration. Nanomanufacturing and Metrology, 8(1), Article 33. https://doi.org/10.1007/s41871-025-00280-x

Image Credits: AI Generated

DOI: 10.1007/s41871-025-00280-x

Keywords: lateral scanning white-light interferometry, SuperPoint, SuperGlue, optical metrology, nanoscale surface topography, image registration, phase-shift error, straightness error, spline interpolation, moving sine wave fitting, seven-step phase shifting, Köhler illumination, microstructure characterization, zero optical path difference, noncontact profilometry, interferogram reconstruction

Cite Scienmag News

Blake Davidson. (November 20, 2026). Deep Learning Registration Corrects Scanning Errors in High-Speed Optical Nanoprofiling. Scienmag. https://scienmag.com/deep-learning-registration-corrects-scanning-errors-in-high-speed-optical-nanoprofiling/

Blake Davidson. "Deep Learning Registration Corrects Scanning Errors in High-Speed Optical Nanoprofiling." Scienmag, 20 November 2026, https://scienmag.com/deep-learning-registration-corrects-scanning-errors-in-high-speed-optical-nanoprofiling/. Accessed 20 November 2026.

Blake Davidson. "Deep Learning Registration Corrects Scanning Errors in High-Speed Optical Nanoprofiling." Scienmag. November 20, 2026. https://scienmag.com/deep-learning-registration-corrects-scanning-errors-in-high-speed-optical-nanoprofiling/

Tags: image registrationinterferogram reconstructionKöhler illuminationlateral scanning white-light interferometrymicrostructure characterizationmoving sine wave fittingnanoscale surface topographynoncontact profilometryOptical metrologyphase-shift errorseven-step phase shiftingSpline interpolationstraightness errorSuperGlueSuperPointzero optical path difference
Share26Tweet16
Previous Post

Researchers Evaluate Interpretable Machine Learning for Cardiac Risk in Hip Surgery

Next Post

Rare ADNP Missense Mutation Alters Chromatin Regulation and Brain Cell Development

Related Posts

Five-Layer Climate Framework Connects Physical Hazards to Real-World Human Impacts
Earth Science

Five-Layer Climate Framework Connects Physical Hazards to Real-World Human Impacts

November 20, 2026
Consensus Level Restriction Curbs Simulated 51 Percent Attacks in Ethereum Classic
Science News

Consensus Level Restriction Curbs Simulated 51 Percent Attacks in Ethereum Classic

November 20, 2026
Inorganic Contaminants in Recycled Plastics Dictate Zeolite Catalyst Deactivation During Cracking
Chemistry

Inorganic Contaminants in Recycled Plastics Dictate Zeolite Catalyst Deactivation During Cracking

November 20, 2026
Self Attention Siamese Neural Networks Boost Accuracy in Case Based Reasoning Retrieval
Science News

Self Attention Siamese Neural Networks Boost Accuracy in Case Based Reasoning Retrieval

November 20, 2026
Griffon Vulture Algorithm Accurately Extracts Seven Nonlinear Parameters in Fuel Cell Simulations
Technology and Engineering

Griffon Vulture Algorithm Accurately Extracts Seven Nonlinear Parameters in Fuel Cell Simulations

November 20, 2026
Simulation-Guided Peptide-Metal Interfaces Advance Precision Metallo-Immunotherapy Design
Science News

Simulation-Guided Peptide-Metal Interfaces Advance Precision Metallo-Immunotherapy Design

November 20, 2026
Next Post
Rare ADNP Missense Mutation Alters Chromatin Regulation and Brain Cell Development

Rare ADNP Missense Mutation Alters Chromatin Regulation and Brain Cell Development

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Five-Layer Climate Framework Connects Physical Hazards to Real-World Human Impacts
  • Consensus Level Restriction Curbs Simulated 51 Percent Attacks in Ethereum Classic
  • Inorganic Contaminants in Recycled Plastics Dictate Zeolite Catalyst Deactivation During Cracking
  • Self Attention Siamese Neural Networks Boost Accuracy in Case Based Reasoning Retrieval

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading