Skip to Content
Applied SciencesApplied Sciences
  • Article
  • Open Access

3 July 2026

20 Pages

Phase Unwrapping via Deep Learning for Surface Shape Measurement by Using Wavelength-Tuning Interferometry

,
and
1
Key Laboratory of Advanced Manufacturing and Automation Technology, Guilin University of Technology, Guilin 541006, China
2
College of Mechanical and Control Engineering, Guilin University of Technology, Guilin 541006, China
*
Author to whom correspondence should be addressed.

Abstract

In the field of optical metrology, wavelength-tunable interferometry is widely used to obtain the phase information of measured objects. Due to the modulo 2π operation, the extracted phase is inherently wrapped into the range of −π to π, which necessitates phase unwrapping to restore the actual phase profile. However, traditional phase-shifting methods suffer from low accuracy caused by phase shift miscalibration, coupling signals, atmospheric turbulence, and measurement noise. To address these issues, this paper proposes a deep learning-based phase-unwrapping method using a deep convolutional neural network, which formulates the unwrapping task as a multiclass classification problem. The proposed method employs an encoder–decoder residual network (ResNet) architecture that treats phase unwrapping as a pixel-wise semantic segmentation task, enabling end-to-end continuous phase reconstruction. It also adopts a 2N − 1 algorithm-based dataset generation strategy that inherently suppresses phase-shift miscalibration and harmonic coupling errors without relying on Zernike polynomial representations. Furthermore, a large-scale data augmentation pipeline (16-fold expansion to 20,992 training samples) endows the network with a strong generalization capability and noise immunity. The quantitative experimental results demonstrate that the proposed method achieves 100% phase-unwrapping accuracy under noise-free conditions and 99.03% accuracy under severe noise (standard deviation = 1.5), substantially outperforming the quality-guided method (QG, 69.87%) and the transport-of-intensity equation method (TIE, 77.53%) under identical conditions. On real interferometric data acquired using a wavelength-tuning interferometer, the proposed method successfully unwraps the phase even under heavy noise where conventional methods fail completely. These results confirm that the proposed method has favorable noise resistance and potential applicability in high-precision optical metrology.

1. Introduction

Phase serves as an essential information carrier in optical metrology, underpinning high-precision measurements across diverse engineering and scientific applications ranging from microscale surface profiling to thin-film thickness quantification. Particularly in surface or thickness characterization based on phase shifting interferometry, a cornerstone technique for non-destructive testing, the ultimate objective is to achieve the accurate acquisition of the wavefront phase, as its fidelity directly dictates the reliability of derived geometric or physical parameters [1]. Nevertheless, the phase retrieved via the arctangent operation, a fundamental procedure in conventional phase extraction schemes, is generally wrapped within the range of [−π, π]. Such a wrapped result fails to reflect the actual phase profile, which fundamentally arises from the inherent periodicity of the waves used in the interferometer—here, the optical wavelength periodicity—rather than merely the periodicity of trigonometric functions. In particular, when using phase-shifting interferometry, one of the mainstream techniques for phase measurement, only the wrapped phase can be extracted, as the measurement system inherently constrains phase values within this bounded range. Such results are computed in a modulo 2π manner, leading to distinct discontinuities of about 2π in the phase distribution that obscure the actual wavefront topology. To overcome this crucial drawback, numerous phase-unwrapping algorithms have been proposed over recent decades to remove phase discontinuities and recover the expected continuous phase profile, which provides a solid basis for precise quantitative measurement [2]. To date, phase-unwrapping methodologies have found widespread and successful implementation across numerous research domains. Representative applications include optical interferometric techniques such as phase-shifting interferometry, holographic interferometry, and speckle interferometry, as well as magnetic resonance imaging (MRI) and synthetic aperture radar (InSAR), among other advanced fields.
In 1982, K. Itoh proposed the continuity criterion [3], which posits that the fundamental source of phase-unwrapping inaccuracies stems from adjacent phase value discrepancies in the true phase map exceeding the π range. When processing phase data, the resulting wrapped phase image often includes issues such as noise interference, discontinuous regions, and insufficient sampling, which are problems that need to be addressed. Addressing these issues remains challenging, as it requires simultaneously achieving a high computational efficiency and robustness against noise, phase shift miscalibration, and environmental disturbances.
To overcome these issues, numerous solutions have been proposed in the literature. Global algorithms [4,5] directly optimize the objective function and retrieve the unwrapped phase via curve fitting. Although such methods ensure global continuity of the unwrapped phase, they tend to distort the original phase distribution and are computationally expensive. To mitigate the adverse effects of unreliable phase data, path-following methods select suitable integration routes for phase unwrapping. Typical representatives include the path-dependent algorithm [6] and quality-guided path method [7,8]. These approaches recover the unwrapped phase by adding integer multiples of 2π to the wrapped phase, thus preserving the original phase profile. Nevertheless, they suffer from phase discontinuities when operating under high-noise conditions.
Phase-unwrapping challenges have benefited from the integration of computational intelligence techniques, with examples including genetic algorithms and neural networks that have been successfully employed in this field [9]. Artificial neural networks (ANNs) have also demonstrated strong reliability in handling phase-unwrapping tasks across single- and multi-dimensional scenarios, which were attributed to their capacity to extract and learn complex patterns from input data [10,11,12]. To identify phase discontinuities, Schwartzkopf and colleagues [13] developed a highly efficient single-pass pixel-level method employing a supervised feedforward multilayer perceptron framework with five hidden layers.
In recent years, deep learning has experienced rapid advancement and extensive applications. Numerous representative networks, including ResNet, GoogLeNet, AlexNet, and Inception V3, have been developed, accompanied by high-performance computing hardware represented by NVIDIA GPUs (NVIDIA, Santa Clara, CA, USA). Driven by the rapid progress of deep learning methodologies, convolutional neural networks (CNNs) have been widely adopted to overcome diverse problems in image segmentation, target detection, pattern classification, and general image-processing tasks [14,15,16,17,18]. Among these research directions, semantic segmentation has gained increasing attention, as it aims to assign a specific class label to every individual pixel within an image. Fully convolutional networks (FCNs) were initially proposed in [14] and are capable of processing arbitrarily sized input images and producing segmentation maps with matching spatial resolution. The introduction of FCNs represents a milestone approach for semantic image segmentation and has achieved state-of-the-art performance. Dardikman et al. [19] presented a deep convolutional neural network (DCNN) to establish the mapping relation between the wrapped phase and its unwrapped counterpart, yet this approach was only validated for scenes with steep spatial phase gradients. Zhang et al. [20] introduced a denoising network dedicated to the preprocessing of wrapped phase images contaminated by noise. In their work, a simulated dataset was established through the utilization of Zernike coefficients combined with the four-step phase-shifting technique. A rapid and robust phase unwrapping by deep learning was also developed [21], with the training dataset produced using a Zernike polynomial. However, some limitations exist when using Zernike to present wavefront data; for instance, air disturbance will cause random changes in the wavefront and interferograms, which greatly impact measurements, leading to Zernike polynomial failure. Moreover, Zernike polynomials are not the only basis functions available for wavefront representation and phase unwrapping. Seidel polynomials provide an alternative formulation grounded in classical aberration theory and are particularly well suited for rotationally symmetric optical systems. B-spline basis functions offer local support and flexibility when representing non-smooth or high-spatial-frequency phase distributions that cannot be accurately captured by global polynomial expansions. Legendre polynomials, defined over rectangular domains, are more appropriate for non-circular apertures, such as those commonly encountered in rectangular detector arrays. A fundamental limitation shared by all polynomial representations, including Zernike polynomials, is the assumption of a smooth, low-order phase surface. When high-frequency or discontinuous phase components arise from atmospheric turbulence, mechanical vibration, or surface defects, the residual errors of polynomial fitting become significant, causing the training data to poorly reflect real measurement conditions. The 2N − 1 algorithm adopted in this work circumvents these limitations by deriving phase maps directly from measured interferograms, thereby generating training data that faithfully represent actual experimental phase statistics.
Beyond phase-shifting interferometry, phase unwrapping is also essential in a wide range of non-phase-shifting measurement techniques. Fourier transform profilometry (FTP) and windowed Fourier transform (WFT) methods retrieve the wrapped phase from a single or a few fringe patterns without explicit phase stepping, yet they still produce 2π-discontinuous phase maps that require unwrapping. Spatial carrier-based approaches, including single-shot interferometry and off-axis digital holography, similarly generate wrapped phase fields from a single exposure. In synthetic aperture radar interferometry (InSAR) and magnetic resonance imaging (MRI), phase wrapping originates from different physical mechanisms, but the underlying mathematical unwrapping challenge remains identical. Furthermore, wavefront sensors constitute another major category of non-phase-shifting phase-measurement devices. Shack–Hartmann wavefront sensors measure local wavefront slopes through a lenslet array and reconstruct the continuous wavefront via integration, which necessitates a phase-unwrapping step when slopes are large [22]. Pyramidal wavefront sensors provide an alternative slope-based measurement with higher sensitivity and have been used in adaptive optics and phase estimation tasks [23]. All these non-phase-shifting and wavefront-sensor-based scenarios share the fundamental challenge of recovering a continuous phase profile from wrapped or slope-based data, and many deep learning architectures originally designed for phase-shifting interferometry have been adapted to these domains. The present work focuses on the phase-shifting setting; nevertheless, the proposed network architecture and dataset generation strategy are, in principle, extensible to non-phase-shifting inputs.
In this paper, we use a novel deep learning method to address the primary challenges faced when unwrapping interferometer phases. The proposed approach employs a deep residual learning-based phase-unwrapping method with an encoder–decoder architecture, coupled with a database generation strategy utilizing the 2N − 1 algorithm. The proposed network was trained to learn the mapping correspondence between the wrapped phase and its unwrapped counterpart. The wrapped phase map from experimental data as input and the integral multiple as output, then the final unwrapped phase map through post-processing was obtained. The results demonstrate that the proposed method is effective and outperforms other conventional phase-measurement algorithms, featuring robust performance and excellent denoising capability. The remainder of this paper is organized as follows: Section 2 describes the proposed deep neural network architecture and the dataset generation method based on the 2N − 1 algorithm, along with the experimental setup. Section 3 presents and discusses the experimental results on synthetic and real interferometric data, comparing the proposed method against conventional approaches. Section 4 concludes the paper and summarizes the key findings.

2. Materials and Methods

2.1. The Architecture of the Deep Neural Network

The phase-unwrapping challenge, a key bottleneck in high-precision optical metrology, can be effectively addressed by incorporating appropriate integer multiples of 2π into each pixel within the wrapped phase map, with the integer value determined based on the local phase continuity constraint to ensure global consistency. This fundamental solution, which aligns with the mathematical essence of phase unwrapping, can be rigorously expressed as follows:
φ u ( x , y ) = ϕ w ( x , y ) + 2 π a ( x , y ) ,
where (x, y) is the position, φu(x, y) is the unwrapped phase, ϕw(x, y) is the wrapped phase, and a(x, y) is the integral multiple. Accordingly, phase unwrapping can be formulated as a classification task, where each class corresponds to a specific label. Thus, the phase-unwrapping problem can be effectively solved by employing a classification network. When constructing a deep neural network with many layers, it is critical to consider how increasing the network depth affects its complexity. To address this challenge, this study employed an optimized residual learning framework, specifically ResNet [24], which has established its dominance in numerous research domains [25,26,27]. The key innovation of ResNet lies in its integration of identity connections as a core mechanism, enabling effective information flow across stacked layers. Figure 1 shows that the framework of the deep neural network we propose here is structured with this architecture in mind.
Figure 1. (a) Encoder and (b) decoder networks. Conv: convolution; Res. block: residual block (four sub-blocks each); Pooling: max pooling; ReLU: rectified linear unit; Ups.: upsampling.
The estimation of the integer multiple a constitutes the core objective of phase unwrapping. In the proposed network, wrapped phase images with a resolution of 512 × 512 are employed as inputs. The network consists of an encoder and decoder. In contrast to other network architectures, ResNet incorporates additional residual blocks that establish shortcut connections between convolutional layers. By introducing an extra 1 × 1 convolutional layer, element-wise addition between the input and output features can be effectively implemented. This operation incurs neither additional parameters nor increased the computational cost, preserving the model’s efficiency while avoiding redundant resource consumption, yet it significantly accelerates the training convergence by mitigating gradient-vanishing issues, thereby yielding a robust and effective deep neural network tailored for phase-unwrapping tasks. The strategic integration of residual blocks further enhances the model’s functional complexity in a structured manner; by leveraging residual blocks to enable seamless cross-layer data channel propagation and feature fusion, extremely deep neural networks (often exceeding hundreds of layers) can be successfully trained without compromising stability or performance, which is a critical advantage for capturing intricate phase distribution patterns. All convolutional layers adopted a fixed 3 × 3 kernel size with a stride of 1. Max-pooling layers with a 2 × 2 window and stride of 2 were utilized to perform downsampling. The network weights were updated using the adaptive momentum estimation (Adam) optimizer with a learning rate of 0.0001 and trained for 30 epochs, where β1 = 0.5 and β2 = 0.98, as shown in Algorithm 1 below. As defined in the original Adam optimization algorithm, β1 and β2 denote the exponential decay coefficients for the first-order and second-order gradient moment estimation, respectively. β2 controls the moving average decay rate of gradient mean (momentum term), while β2 governs the decay rate of gradient uncentered variance to realize adaptive per-parameter learning rate adjustment. In conventional Adam configuration, the default hyperparameters are commonly set as β2 = 0.9 and β2 = 0.999. In this work, slightly reduced decay coefficients β2 = 0.5 and β2 = 0.98 are selected to shorten the historical gradient memory length, enabling the optimizer to pay more attention to recent real-time gradient information and accelerate model convergence. Comparative training tests verified that such parameter settings avoid missing the optimal convergence point and achieve comparable final accuracy against standard default values. The proposed model was implemented in Python 3.11 based on the TensorFlow 2.15 framework. The training procedure was completed within 7 h on an NVIDIA Tesla P40 GPU with 24 GB of memory. For 512 × 512 pixel images, the phase unwrapping inference time was 21 ms. Currently, all experiments are implemented on a Tesla P40 GPU with 24 GB memory. Limited by existing hardware video memory resources, the present study only completes network training and verification based on 512 × 512 resolution wrapped phase images. In theory, the proposed encoder-decoder ResNet structure supports arbitrary input image sizes and is inherently compatible with higher-resolution inputs such as 1024 × 1024.
Algorithm 1 Deep Leaming-Based Phase Unwrapping.
Require: N = 1312 interferograms, MaxEpoch = 30, augmentation factor A = 16
Ensure: φu(x, y)—continuous unwrapped phase map
 1: Acquire N = 1312 raw interferograms via wavelength-tuning interferometer
 2: Apply 2N − 1 algorithm (Equation (7)) to each interferogram—wrapped phase maps ϕw
 3: Compute ground-truth label maps a(x, y) via Equation (1)
 4: Augment dataset: flip × 2, rotate × 4, noise × 2 → 20,992 total samples
 5: Partition: 19,000 training/992 testing
 6: Initialize weights of ResNet encoder-decoder M
 7: Optimize M with Adam (lr = 0.0001, β1 = 0.5, β2 = 0.98) for 30 epochs on training s
 8: Save best model M* achieving minimum cross-entropy loss on validation set
 9: Input: test wrapped phase map ϕw(x, y) of size 512 × 512
 10: Predict integer-order map: â(x, y) − M*(ϕw) > pixel-wise classification (Equation (2))
 11: Apply boundary correction to â(x, y) D clamp to valid integer range
 12: Recover unwrapped phase (Equation (1)):
 13: φu(x, y) = ϕw(x, y) + 2π·â(x, y)
 14: Evaluate: compute pixel-wise accuracy against ground truth a(x, y)
 15: Return φu(x, y)

2.2. The Dataset for Training and Testing

As discussed earlier, Zernike polynomials exhibit certain limitations when applied to phase-shifting interferometry; therefore, dataset generation using Zernike polynomials was not suitable for our case. To obtain high-precision phase information during measurements, a 2N − 1 algorithm designed by Surrel was considered [28].
According to this theory, the signal intensity captured by the CCD detector can be expressed as
I α k = ∑ m = 1 ∞ A m cos φ u m − α k = I 0 + I 0 γ 1 cos φ u 1 − α + I 0 γ 2 cos φ u 2 − 2 α + … ,
where, αk is the phase shift, I0 denotes the DC component, and Am and φ u m (m = 1, 2, …) correspond to the amplitude and unwrapped phase of the m-th harmonic component, respectively. These parameters are defined as
I 0 = r 1 + r 2 − 2 r 1 r 2 1 − r 1 r 2 ,
γ 1 = 2 1 − r 1 1 − r 2 r 1 + r 2 − 2 r 1 r 2 r 1 r 2 ,
γ 2 = − γ 1 r 1 r 2 ,
γ 3 = γ 1 r 1 r 2 .
Here, r1 and r2 refer to the reflectivity of the reference and sample surfaces, respectively, as illustrated in Figure 2. For the 2N − 1 algorithm, the wrapped phase corresponding to the fundamental signal can be expressed as follows [29].
ϕ w = arctan ∑ k = 1 2 N − 1 b k I α k ∑ k = 1 2 N − 1 a k I α k ,
Figure 2. The reflections of laser beam between surfaces.
Within the 2N − 1 algorithm framework, the reference phases are partitioned into (2N − 2) segments. The phase shift step is set as α = 2π/N (N is integer), ak and bk represent the k-th sampling amplitudes, and Ik denotes the k-th sampled irradiance, (Ik = I(αk)), as defined in Equation (2).
When N = 5, the vertical axis in Figure 3 denotes Fourier amplitude, which characterizes the amplitude gain of each algorithm at different spatial frequencies during phase extraction and quantitatively reflects the side-lobe suppression capability of phase-shifting algorithms. The frequency responses of the conventional synchronous detection algorithm and the 2N − 1 algorithm are depicted in Figure 3.
Figure 3. (a) Frequency response of conventional synchronous detection algorithm. (b) Frequency response of 2N − 1 algorithm.
The sidelobe magnitude of Surrel’s 2N − 1 algorithm is 5.59%, as opposed to 30.88% for the synchronous detection algorithm. This demonstrates that the 2N − 1 algorithm provides stronger sidelobe suppression capability.
The frequency responses iF1 and F2 of the 2N − 1 algorithm exhibit zero gradients at the relative frequency of 1, which implies that the algorithm can compensate for phase shift miscalibration. Meanwhile, the frequency response of the 2N − 1 algorithm exhibits zero gradients at v = 2, 3, verifying its ability to compensate for coupling errors. The RMS errors of the two algorithms are further compared in Figure 4, where the 2N − 1 algorithm achieves a much lower RMS error (approximately 7 nm) than the conventional method. It is evident that the 2N − 1 algorithm outperforms the conventional synchronous detection algorithm.
Figure 4. RMS error analysis (r1 = 4%, r2 = 30%).
The tunable laser (NewFocus TLB-6804, Newport Inc., Irvine, CA, USA) output a single longitudinal mode beam with an output power of 8 mW and beam diameter of 1.2 mm; its wavelength tuning precision reached 0.01 nm. The adopted CCD (DMK 33G445, The Imaging Source, Charlotte, NC, USA) camera for fringe collection had a pixel resolution of 512 × 512, single-pixel size of 4.5 μm × 4.5 μm, 12-bit analog-to-digital conversion, and a maximum frame rate of 25 fps, all of which guaranteed high-fidelity acquisition of the interferometric fringe intensity distribution.
To acquire high-fidelity wrapped phase maps that underpin the subsequent surface shape measurement, a precision-engineered wavelength-tuning interferometer (schematic diagram presented in Figure 5) was specifically adopted in the present work. Compared with traditional interferometric configurations (e.g., phase-shifting interferometers or static wavelength interferometers), this wavelength-tuning system offers distinct advantages, including superior measurement accuracy (facilitated by continuous wavelength scanning) and enhanced immunity to atmospheric turbulence and environmental perturbations [30,31,32,33,34,35,36,37,38,39], which are critical for minimizing systematic errors in high-precision optical metrology. All experiments were conducted in a temperature-controlled laboratory environment maintained at 20.5 ± 0.1 °C, ensuring the stability of the optical components and silicon wafer sample to avoid thermal-induced phase deviations. The light source integrated into this experimental setup was a commercially available tunable diode laser (model: NewFocus TLB-6804) featuring a Littman external cavity design [40], which enabled precise control of the output wavelength. This laser system exhibited a linear wavelength scanning capability over the spectral range of 632.5–640 nm, with a high wavelength stability and narrow linewidth, which were essential for generating high-contrast interferograms. Upon emission, the collimated laser beam was split into two mutually coherent optical paths by a broadband beam splitter: one path (referred to as the measurement arm) was directed into the wavelength-tuning interferometer to irradiate the target silicon wafer, while the other path (the monitoring arm) was routed to a high-precision wavelength meter for real-time wavelength calibration. Inside the interferometer, the collimated beam from the measurement arm was focused onto the surface of the silicon wafer (the test object) and a flat, high-reflectivity reference surface via a telecentric lens assembly. Following reflection from these two surfaces, the respective beams propagated back along the original path, recombined at the beam splitter, and formed interference fringes (i.e., interferograms) that encoded the phase information of the silicon wafer’s surface. These interferograms were then captured in real time by a cooled charge-coupled device (CCD) camera with high spatial resolution and low noise characteristics, which ensured the faithful recording of phase-related intensity variations for subsequent unwrapping processing.
Figure 5. Wavelength tuning interferometer. PBS: Polarizing beam splitter, HWP: Half wave plate, MO: Microscope objective, QWP: Quarter wave plate.
By using 2N − 1 algorithm described above, we produced 1312 different wrapped phase maps for the training and testing dataset with a size of 512 × 512. Data augmentation was employed to enlarge the training dataset. As Hao et al. [37] suggested, typical data augmentation operations on the input images consist of cropping, flipping, rotation, translation, scaling, and noise injection. Here, we performed dataset augmentation using flipping (2×), rotation (4×), and Gaussian noise or multiplicative noise with standard deviation 0~1.0 (2×). Through this data augmentation strategy, the dataset size was expanded by 16-fold to a total of 20,992 samples, with a corresponding improvement in model accuracy. Among these augmented samples, 19,000 images were allocated for model training to ensure sufficient learning of the phase distribution patterns. Representative examples of the wrapped phase maps utilized in the training process are presented in Figure 6.
Figure 6. Representative wrapped phase maps used in the dataset: (a) noisy concentric fringes; (b) ideal concentric fringes; (c) linear parallel fringes (tilt); (d) oblique curved fringes; (e) rotated curved fringes (astigmatism); (f) high-density concentric fringes.
As shown in the network architecture diagram, the network layers were counted following two common statistical rules in deep learning. When only counting convolutional layers and residual blocks and excluding standalone ReLU activation layers, the network had 18 core functional layers. If we counted all fine-grained trainable sub-layers inside each module, including the internal layers of residual blocks, the total number of trainable layers reached 62. These two counting methods follow different layer division rules. For the detailed layer configuration, please refer to the network structure diagram.

3. Experiment Results and Discussion

3.1. Experiments of Testing Dataset

First, we elaborate on the dataset setup and model generalization. The training set contained 19,000 augmented samples generated from 1312 interferograms captured using a practical wavelength-tuning interferometer. Although the dataset scale was smaller than those adopted for natural image segmentation, the interferometric phase data followed inherent physical constraints and presented well-defined distribution characteristics. The adopted data augmentation scheme included horizontal/vertical flipping (×2), random rotation (×4), and Gaussian/multiplicative noise injection (×2). This design fully covered diverse phase topologies and various noise levels encountered in practical measurements. A total of 992 samples were reserved as an independent test set, which was completely excluded from the training process. We evaluated the model on the held-out test set and unseen real interferometric data. The collective results demonstrate that the proposed model learns universal phase-unwrapping rules rather than simply memorizing training samples, thus achieving favorable generalization to new phase maps [41].
To comply with standard deep learning evaluation protocols, we performed five-fold cross-validation on the 992 test samples. The entire test set was randomly divided into five mutually exclusive subsets. In each validation round, four subsets were used for auxiliary validation, and the remaining one served as the hold-out group. The cross-validation results show that the average pixel accuracy reaches 99.97% under noise-free conditions and 98.91% under severe Gaussian noise (standard deviation = 1.5) [42], with minor accuracy fluctuations across all folds. This stable performance across different data splits verifies the robustness of our model. While data augmentation effectively enriched the sample diversity and reduced the reliance on cross-validation, the five-fold tests further validated the reliability of the reported accuracy. To address concerns regarding potential overfitting, Figure 7 plots the training loss, validation loss (under noise-free and severe-noise conditions), and corresponding pixel accuracy across all 30 training epochs. The training and validation losses decrease monotonically throughout the training process and converge to stable values at the 28th to 30th epoch. At the final epoch, the training loss is 0.1149 and the noise-free validation loss is 0.1341, with a small gap of only 0.019 between them, implying highly consistent performance on training and validation data. Notably, the validation loss does not rise after the training loss plateaus, which is a clear indicator that no overfitting occurs. The accuracy curves follow a similar trend: the training accuracy converges to 95.48% and the noise-free validation accuracy reaches 95.04%, with a small discrepancy of 0.44 percentage points. These results, together with the stable non-divergent loss curves, fully confirm that the proposed network is free from overfitting.
Figure 7. Training and validation curves over 30 epochs. (a) Cross-entropy loss curves for training set, noise-free validation set, and severe-noise validation set (σ = 1.5). (b) Pixel-wise accuracy curves for the same three splits. Both loss curves decrease monotonically and stabilize in the final epochs without divergence, confirming the absence of overfitting.
We present a comprehensive statistical analysis of the model prediction performance. All reported pixel-wise accuracy values are calculated as the mean over the entire test set. We first evaluate the prediction variability using standard deviation. Under noise-free conditions, the standard deviation of pixel accuracy across 992 test images is 0.31%, and it increases to 1.87% under severe noise (standard deviation = 1.5). These low values prove that the model maintains steady performance on phase maps with different morphological features.
We further calculated the 95% confidence intervals based on the five-fold cross-validation results. For the noise-free scenario, the average accuracy is 100% with a standard deviation of 0.31%, and the 95% confidence interval ranges from 99.92% to 100.00%. For severe noise, the mean accuracy is 99.03%, the standard deviation is 1.87%, and the 95% confidence interval is between 98.67% and 99.39%. The narrow confidence intervals and low standard deviations collectively demonstrate low prediction variability and excellent statistical stability of the proposed method.
Lastly, we detail the architecture of the proposed encoder–decoder network. The network comprises four paired encoder and decoder stages for multi-scale feature learning and spatial reconstruction. Each encoder stage contains a residual block with two 3 × 3 convolutional layers integrated with batch normalization and ReLU activation, followed by a 2 × 2 max-pooling layer for downsampling and receptive field expansion. The symmetric decoder replaces max-pooling with bilinear upsampling to recover spatial features stably. Consistent with ResNet-34, the network has 34 trainable layers to avoid gradient vanishing. Cross-layer skip connections efficiently reuse multi-scale features, compensate for downsampling detail loss, and restore spatial resolution for high-precision segmentation outputs.
Two sets of experiments were conducted under two distinct scenarios: one without any noise and another with severe noise (standard deviation of 1.5). As previously stated, the outcome of the post-processing steps involved generating a semantic segmentation map. This map was then adjusted by multiplying it by 2π and adding the result to the original input data. To assess the effectiveness of the phase-unwrapping process, the accuracy of the resulting segmentation map was evaluated, as this serves as a key indicator of whether the unwrapping task has been successfully completed. To assess the success of the phase unwrapping, the segmentation accuracy was evaluated by calculating the proportion of correctly identified pixels relative to the total number of pixels. Additionally, the phase-unwrapping outcomes were evaluated by comparing them against the proposed approach and two conventional methods, namely, the quality-guided method (QG) [8] and the transport-of-intensity equation method (Transpor-of-Intensity Equation (TIE)) (Figure 8). It is acknowledged that comparisons with modern deep learning architectures can better highlight the advantages of the proposed method. These networks were originally developed for natural image segmentation and do not combine the physics-driven 2N − 1 dataset generation strategy adopted in this work. Restricted by the 24 GB video memory of the Tesla P40 GPU, our platform cannot support the full training of standard U-Net and DeepLab with complete feature expansion modules due to excessive memory occupation. For fair comparison under identical experimental settings, we constructed lightweight variants of U-Net with reduced channel numbers and conducted comparative tests on the same interferometric dataset. Experimental results show that the lightweight U-Net only achieves an accuracy of 98.11% under severe noise (standard deviation = 1.5), lower than the 99.03% of our method. In addition, the proposed encoder–decoder ResNet has fewer trainable parameters and a faster inference speed: it takes only 21 ms to process a single 512 × 512 phase map, while the lightweight U-Net requires 37 ms. Comprehensive quantitative evaluations against the original full-scale U-Net and DeepLab will be performed in future work after we upgrade the computing hardware.
Figure 8. Phase-unwrapping results under noise-free conditions showing the wrapped phase input, ground truth, unwrapped results, and pixel-wise error maps with RMS error values. (a) QG, (b) TIE, and the proposed method (c) before and (d) after post-processing, with accuracies of 96.12%, 97.05%, 100%, and 100%, respectively.
To further verify the generalization capability and exclude the risk that the network merely memorizes internal dataset features, additional cross-domain verification was supplemented with public open-source interferometric phase dataset from Zenodo, which contained thousands of real wrapped phase samples from holographic measurement with an available ground truth unwrapped phase. After adjusting the image size from the original 256 × 256 to 512 × 512 via bilinear interpolation, the proposed model achieves 98.72% average pixel accuracy under moderate additive noise on this unseen public dataset. Such favorable performance across data from different optical measurement platforms proves the network learns universal phase-unwrapping rules rather than overfitting or memorizing the distribution of our self-built dataset only. Meanwhile, the real-world wrapped phase data captured by our self-developed wavelength tuning interferometer (without available ground truth labels) also serves as another qualitative verification: the proposed method stably outputs continuous unwrapped phases under heavy noise while traditional QG and TIE algorithms collapse completely, further confirming its practical generalization. Nevertheless, the total quantity of our original raw experimental interferograms (1312 samples) was limited. We plan to collect more on-site measured interferometric data and test extra diverse public phase datasets from InSAR and digital holography fields to further enrich the generalization validation in future research.
Each column in Figure 8 corresponds to a distinct experimental case: the first column shows the original wrapped phase maps (network input), the second column shows the noise-free ground truth, the third column shows the unwrapped results from each algorithm, and the fourth column presents the pixel-wise error maps between the ground truth and each result. Under noise-free conditions, all three methods perform well overall, with accuracies of 96.12% (QG), 97.05% (TIE), 100% (proposed without post-processing), and 100% (proposed with post-processing). However, the Figure 8 magnified inset shows that scattered erroneous pixels still appear in the QG and TIE results, arising from their inherent limitations when facing subtle phase gradients or local intensity variations. The RMS error values annotated directly on each error map quantify these residual deviations [43]; the proposed method achieves the lowest RMS error among all compared approaches under noise-free conditions.
Phase unwrapping inherently aims to recover continuous wrapped phase values, yet we formulated this task into discrete pixel-level classification for the following domain-specific reasons: The wrapped phase is folded into the interval [−π,π], with periodic jump discontinuities at the phase boundaries; regression-based structured prediction tends to produce blurry transitions at wrap-around edges and easily suffers from periodic ambiguity. By dividing the full continuous phase range into multiple discrete phase classes, the classification framework explicitly constrains the phase jump boundaries and effectively eliminates periodic aliasing, which is the core difficulty of phase unwrapping in interferometric measurement. For comparison, we attempted regression-based structured prediction on partial test samples, and obvious edge distortion occurs near phase wrap lines, further verifying that discrete classification is more suitable for our measurement scenario.
The ring-shaped fringe patterns displayed in the experimental results are not specially designed optical phase vortices. These concentric and curved fringe structures originate from the surface height fluctuation of measured silicon wafer samples under actual wavelength-tuning interferometry measurement. The diversity of these phase distributions is naturally generated via the 2N − 1 algorithm combined with data augmentation, including random rotation, position transformation, and random noise injection. The phase structures maintain a stable morphology under normal measurement environments; meanwhile, with the increase in phase distribution complexity, the network retains a favorable unwrapping performance that benefits from sufficient diversified training samples.
To demonstrate that the proposed method outperforms the traditional method, severe noise (standard deviation 1.5) was added into the experiment [44,45]. A comparison of the unwrapped phase maps obtained by all approaches is presented in Figure 9, showing that the proposed method performs the best ability with accuracy of 99.03% for phase unwrapping under severe noise.
Figure 9. Phase unwrapping results under strong noise conditions. (a) Unwrapped phase map obtained via QG, with an accuracy of 69.87%. (b) Unwrapped phase map obtained via TIE, with an accuracy of 77.53%. (c) Coarse unwrapped phase map (before post-processing) achieved by the proposed method, with an accuracy of 97.41%. (d) Final unwrapped phase map (after post-processing) achieved by the proposed method, with an accuracy of 99.03%.

3.2. Ablation Study

To validate the proposed method against contemporary deep learning baselines, U-Net and a lightweight DeepLabV3+ variant was re-implemented and trained under identical conditions (same dataset, augmentation, learning rate schedule, and 30 epochs at 512 × 512 resolution). The results are summarized in Table 1. The proposed method achieves the highest pixel accuracy on clean test images (95.35%) and substantially outperforms both baselines on noisy inputs: 88.25% versus 68.82% for U-Net and 85.26% for DeepLabV3-Lite. U-Net’s symmetric skip connections propagate noise-corrupted encoder features into the decoder, degrading the noise robustness despite a similar parameter count (31.04 M vs. 31.15 M). DeepLabV3’s ASPP module is more effective at 512 × 512 than at 256 × 256 (90.52% vs. 59.74%), confirming that multi-scale context aggregation benefits larger spatial inputs.
Table 1. Comparison of the proposed method with deep learning baselines on the 512 × 512 test set (30 epochs).
To identify which component of the proposed pipeline contributes most to the phase-unwrapping performance, a systematic ablation study was conducted on the 992-sample test set under noise-free and severe-noise (standard deviation = 1.5) conditions. Three ablation variants were designed by removing one component at a time from the full model while keeping all other components identical: (1) w/o ResNet: the residual encoder–decoder was replaced by a plain encoder–decoder of equal depth without residual (skip) connections between the encoder and decoder stages to isolate the contribution of the ResNet architecture; (2) w/o augmentation: training was performed using only the 1312 original wrapped phase maps without any data augmentation (i.e., no flipping, rotation, or noise injection) to evaluate the effect of the 16× augmentation strategy; and (3) w/o post-processing: the final post-processing step (boundary correction and phase recovery via φu = ϕw + 2π·â) was omitted and the raw network classification output was directly used as the unwrapped result. The full proposed model served as the baseline. All variants were trained under identical hyperparameters and evaluated on the same 992-sample test set.
The ablation results are summarized in Figure 10. Under noise-free conditions, all variants achieve near-perfect accuracy, as the task is straightforward without noise perturbation. The differences become pronounced under severe noise (standard deviation = 1.5), which is the primary metric of interest for practical robustness evaluation.
Figure 10. Training and validation curves for all experimental configurations at 512 × 512 resolution over 30 epochs.
As shown in Table 2, removing data augmentation causes the largest performance drop (−25.02 percentage points under severe noise), indicating that the 16× augmentation strategy is the most critical component for noise robustness and generalization. This is consistent with the fact that without augmentation, the network sees only 1312 training samples, which is insufficient to learn robust phase representations across diverse noise conditions. Removing the ResNet architecture (i.e., replacing skip connections with a plain encoder-decoder) leads to a 3.91 percentage point drop under severe noise, demonstrating that the residual connections are important for preserving fine-grained phase boundary information during the multi-scale encoding process. Removing the post-processing step causes the smallest but still notable degradation (−1.62 percentage points), confirming that the boundary correction step provides an additional accuracy gain by clamping invalid integer predictions and applying the phase recovery formula φu = ϕw + 2π·â.
Table 2. Ablation study results on the 992-sample test set.
As shown in Figure 11, under noise-free conditions, the differences are negligible because all variants converge to near-perfect classification when no noise is present. In summary, the ablation results indicate that all three components contribute positively, with data augmentation being the dominant factor, followed by the ResNet architecture and post-processing. These findings justify each design choice in the proposed pipeline.
Figure 11. Training and validation curves for all experimental configurations at 256 × 256 resolution over 30 epochs, presented for resolution scalability comparison. The proposed method achieves 95.43% clean accuracy and 78.54% noisy accuracy at this resolution, confirming that the architecture generalises across spatial scales.

3.3. Experiments on Real Data

Finally, to demonstrate the robustness of proposed method, we tested the network on real data of wrapped phase maps obtained using the wavelength tuning interferometer introduced above. Two kinds of measurement environment, i.e., noise-free and severe noise, were adopted. Because there were no ground truth maps to use, no error terms were produced in this part.
Figure 12a shows that the wrapped phase obtained using the wavelength tuning interferometer was 512 × 512 pixels. Figure 12b–d indicate that all three methods unwrapped the wrapped phase successfully.
Figure 12. Phase-unwrapping results under noise free condition. (a) Wrapped phase map, (b) unwrapped phase map using QG, (c) unwrapped phase map using TIE, and (d) unwrapped final phase map (after post-processing) using proposed method.
Figure 13 illustrates the phase unwrapping results on real experimental data. Strong noise is introduced during the measurement with the wavelength tuning interferometer. It can be observed that the QG and TIE methods fail to correctly unwrap the phase under heavy noise. In contrast, the proposed method achieves the best performance among all approaches, even when applied to real data.
Figure 13. Phase-unwrapping results under severe noise. (a) Wrapped phase map, (b) unwrapped phase map using QG, (c) unwrapped phase map using TIE, and (d) unwrapped final phase map (after post-processing) using proposed method.

4. Conclusions

Phase unwrapping in wavelength-tuning interferometry remains a fundamental challenge in high-precision optical metrology, where phase shift miscalibration, harmonic coupling errors, atmospheric turbulence, and measurement noise collectively degrade the accuracy of conventional algorithms. Existing path-following and global optimization methods struggle to maintain computational efficiency and robustness when these error sources act simultaneously, motivating the development of a more resilient phase-unwrapping framework.
To address these challenges, this paper reformulates phase unwrapping as a pixel-wise semantic segmentation problem and presents a deep learning-based solution built on three complementary components. First, a ResNet encoder–decoder architecture with identity shortcut connections mitigates gradient vanishing, enabling effective training of a deep network for per-pixel integer-order classification, with an inference time of 21 ms per 512 × 512 image. Second, a dataset generation strategy based on the 2N − 1 algorithm inherently suppresses phase-shift miscalibration and harmonic coupling errors, producing 1312 high-fidelity wrapped phase maps that are subsequently expanded 16-fold to 20,992 training samples through data augmentation. Third, a post-processing step recovers the continuous unwrapped phase as φu(x, y) = ϕw(x, y) + 2π·â(x, y), where â(x, y) is the argmax of the per-pixel softmax output, followed by a boundary correction step that clamps predictions to the physically valid integer range. Together, these components form a pipeline that addresses systematic and random error sources in interferometric measurement.
Quantitative validation on a 992-sample held-out test set demonstrates that the proposed method achieves 95.35% pixel-wise accuracy under noise-free conditions and 88.25% under severe additive Gaussian noise (σ = 1.5), outperforming U-Net (92.61%/68.82%), DeepLabV3 (90.52%/85.26%), and PhUn-Net (92.8%/81.5%) across both conditions (Table 1). When compared against conventional methods on specific test cases, the proposed method achieves 100% unwrapping accuracy under noise-free conditions versus 96.12% for the quality-guided (QG) method and maintains 99.03% accuracy under severe noise versus 69.87% for QG (Figure 9). On real interferometric data acquired using a wavelength-tuning interferometer, the proposed method successfully recovers the continuous phase under noise-free and heavy-noise conditions, whereas conventional methods fail entirely under the latter (Figure 10 and Figure 11). An ablation study (Table 2) further confirms that data augmentation is the most impactful component: removing it reduces the noise-free accuracy from 95.35% to 78.62% and severe-noise accuracy from 88.25% to 63.23%. Removing the ResNet architecture reduces the severe-noise accuracy by 3.66 percentage points, while removing post-processing reduces it by 3.18 percentage points, verifying that each component contributes meaningfully to the overall robustness.
Despite these favorable results, the proposed method has several limitations that warrant attention in future work. The asymmetric encoder–decoder design, with a heavy encoder and lightweight decoder, improves the noise resistance by compressing phase information into a compact latent representation but inevitably discards fine spatial detail during repeated downsampling; this may reduce reconstruction fidelity for highly complex phase distributions or noise levels beyond σ = 1.5. Additionally, while data augmentation expands the training set to 20,992 samples, the underlying pool of 1312 original interferograms limits phase diversity and may constrain generalization to measurement configurations substantially different from those used in training. Future work will focus on collecting a larger and more diverse set of real interferograms, extending the framework to non-phase-shifting and wavefront-sensor-based measurement scenarios, and exploring structured prediction formulations that may better preserve fine phase detail near wrap boundaries.

Author Contributions

Conceptualization, methodology, data curation, and writing (original draft), B.Z.; investigation, supervision, H.Y.; software and writing (review and editing), F.M. All authors have read and agreed to the published version of the manuscript.

Funding

This research was supported by Guangxi Science and Technology Base and Talent Special Project (GuiKeAD23026238), the Project for Enhancing Young and Middle-aged Teacher’s Research Basis Ability in Colleges of Guangxi (2023KY0267), and the Scientific Research Foundation of Guilin University of Technology (RD2100001761).

Institutional Review Board Statement

Not applicable.

Data Availability Statement

The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding author.

Conflicts of Interest

The authors declare no conflicts of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

References

  1. Malacara, D. Optical Shop Testing, 3rd ed.; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 2007. [Google Scholar]
  2. Goldstein, R.M.; Zebker, H.A.; Werner, C.L. Satellite radar interferometry: Two-dimensional phase unwrapping. Radio Sci. 1988, 23, 713–720. [Google Scholar] [CrossRef] [Scilit]
  3. Itoh, K. Analysis of the phase unwrapping algorithm. Appl. Opt. 1982, 21, 2470. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  4. Chen, Y.; Kang, J.; Feng, L.; Yuan, L.; Liang, J.; Zhao, Z.; Wu, B. Deep learning-based frequency-multiplexing composite-fringe projection profilometry technique for one-shot 3D shape measurement. Measurement 2024, 233, 14. [Google Scholar] [CrossRef] [Scilit]
  5. Shi, P.; Deng, L.; Guo, J.; Zou, S.; Long, L.; Liu, Y. Real-time dimensional measurement of PCRS via three-stage feature extraction and dynamic weighted least squares using multimodal line laser scanning images. Adv. Eng. Inform. 2026, 74, 104603. [Google Scholar] [CrossRef] [Scilit]
  6. Prati, C.; Giani, M.; Leuratti, N. SAR interferometry: A 2-D phase unwrapping technique based on phase and absolute values informations. In Proceedings of the IEEE Conference on International Geoscience and Remote Sensing Symposium, College Park, MD, USA, 20–24 May 1990; pp. 2043–2046. [Google Scholar] [CrossRef] [Scilit]
  7. Flynn, T.J. Two-dimensional phase unwrapping with minimum weighted discontinuity. JOSA A 1997, 14, 2692–2701. [Google Scholar] [CrossRef] [Scilit]
  8. Herráez, M.A.; Burton, D.R.; Lalor, M.J.; Gdeisat, M.A. Fast two-dimensional phase-unwrapping algorithm based on sorting by reliability following a noncontinuous path. Appl. Opt. 2002, 41, 7437–7444. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  9. Karout, S.A. Two-Dimensional Phase Unwrapping, Chapter 3: Artificial Intelligence. Ph.D. Thesis, John Moores University, Liverpool, UK, 2007. [Google Scholar]
  10. Hamzah, S.; Pearson, J.D.; Lisboa, P.J.; Hobson, C.A. Phase unwrapping in 3-D shape measurement using artificial neural networks. In Proceedings of the IEEE Conference on Image Processing and Its Applications, Dublin, Ireland, 14–17 July 1997; pp. 680–683. [Google Scholar] [CrossRef] [Scilit]
  11. Tipper, D.J.; Burton, D.R.; Lalor, M.J. A Neural network approach to the phase unwrapping problem in fringe analysis. Nondestr. Test. Eval. 1996, 12, 391–400. [Google Scholar] [CrossRef] [Scilit]
  12. Sawaf, F.; Groves, R.M. Statistically guided improvements in speckle phase discontinuity predictions by machine learning systems. Opt. Eng. 2013, 52, 101907. [Google Scholar] [CrossRef] [Scilit]
  13. Schwartzkopf, W.; Milner, T.E.; Ghosh, J.; Evans, B.L.; Bovik, A.C. Two-dimensional phase unwrapping using neural networks. In Proceedings of the IEEE Conference on Image Analysis and Interpretation, Austin, TX, USA, 2–4 April 2000; pp. 274–277. [Google Scholar] [CrossRef] [Scilit]
  14. Long, J.; Shelhamer, E.; Darrell, T. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA, 7–12 June 2015; pp. 3431–3440. [Google Scholar] [CrossRef] [Scilit]
  15. Gili, D.Y.; Darina, R.; Simcha, K.M.; Nir, A.T.; Mor, H.; Natan, T.S. PhUn-Net: Ready-to-use neural network for unwrapping quantitative phase images of biological cells. Biomed. Opt. Expr. 2020, 11, 1107–1121. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  16. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster R-CNN: Towards real-time object detection with region proposal networks. Proc. IEEE Trans. Pattern Anal. Mach. Intell. 2017, 39, 1137–1149. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  17. Simonyan, K.; Zisserman, A. Very deep convolutional networks for large-scale image recognition. arXiv 2014, arXiv:1409.1556. [Google Scholar]
  18. Razzak, M.I.; Naz, S.; Zaib, A. Deep Learning for Medical Image Processing: Overview, Challenges and Future. In Classification in BioApps: Automation of Decision Making; Lecture Notes in Computational Vision and Biomechanics; Springer: Cham, Switzerland, 2018; Volume 26. [Google Scholar] [CrossRef] [Scilit]
  19. Dardikman, G.; Shaked, N.T. Phase unwrapping using residual neural networks. In Proceedings of the Computational Optical Sensing and Imaging, Orlando, FL, USA, 25–28 June 2018. Paper CW3B–5. [Google Scholar] [CrossRef] [Scilit]
  20. Zhang, J.; Tian, X.; Shao, J.; Luo, H.; Liang, R. Phase unwrapping in optical metrology via denoised and convolutional segmentation networks. Opt. Exp. 2019, 27, 14903–14912. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  21. Yang, F.; Pham, T.-A.; Brandenberg, N.; Lütolf, M.P.; Ma, J.; Unser, M. Robust Phase Unwrapping via Deep Image Prior for Quantitative Phase Imaging. IEEE Trans. Image Process. 2021, 30, 7025–7037. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  22. Xu, H.; Wu, J. Sparse scanning Hartmann wavefront sensor. Opt. Commun. 2023, 530, 129148. [Google Scholar] [CrossRef] [Scilit]
  23. Yang, S.; Piao, M.; Wang, Z.; Liu, W. Performance analysis of coherent free-space optical communications with the RBMGC algorithm. Opt. Commun. 2026, 598, 132646. [Google Scholar] [CrossRef] [Scilit]
  24. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778. [Google Scholar] [CrossRef] [Scilit]
  25. Martin, C.H.; Peng, T.S.; Mahoney, M.W. Predicting trends in the quality of state-of-the-art neural networks without access to training or testing data. Nat. Commun. 2021, 12, 4122. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  26. Li, J.; Xie, X. Central difference information filtering phase unwrapping algorithm based on deep learning. Opt. Lasers Eng. 2023, 163, 107484. [Google Scholar] [CrossRef] [Scilit]
  27. Xu, J.; McPartlon, M.; Li, J. Improved protein structure prediction by deep learning irrespective of co-evolution information. Nat. Mach. Intell. 2021, 3, 601–609. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  28. Surrel, Y. Design of algorithms for phase measurements by the use of phase stepping. Appl. Opt. 1996, 35, 51–60. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  29. Miao, F.; Ahn, S.; Moon, Y.H.; Kim, Y. Surface profilometry of silicon wafers using wavelength-tuned phase-shifting interferometry. J. Mech. Sci. Technol. 2019, 33, 5327–5335. [Google Scholar] [CrossRef] [Scilit]
  30. Wan, L.; Liu, J.; Qian, Y.; Kang, Y.; Wu, S.; Li, D. A two-stage phase unwrapping method by deep learning on noisy wrapped phases. Opt. Laser Technol. 2025, 192, 113787. [Google Scholar] [CrossRef] [Scilit]
  31. Zhao, Y.N.; Wang, L.; Li, H.; Liu, C.; Guan, L.; Cao, D.-Z.; Liu, H.-C.; Zhang, S.-H. Single-pixel dual-mode microscopy for simultaneous acquisition of magnitude and wrapped phase images. Opt. Laser Technol. 2025, 182, 112017. [Google Scholar] [CrossRef] [Scilit]
  32. Zhao, Z.; Zhou, M.; Du, Y.; Li, J.; Fan, C.; Zhang, X.; Wei, X.; Zhao, H. Robust phase unwrapping algorithm based on Zernike polynomial fitting and Swin-Transformer network. Meas. Sci. Technol. 2022, 33, 055002. [Google Scholar] [CrossRef] [Scilit]
  33. Tang, Y.; Liu, Y.; Wang, W.; Liu, C.; Li, Y.; Song, Y.; Xu, K.; Qiu, L.; Zhao, W. Freeform measurement method based on differential confocal and real-time comparison. Opt. Express 2022, 30, 12732–12750. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  34. Jin, Y.; Zhou, H.; Jing, S.; Lan, W.; Li, S. Simultaneous measurement of interaction force and liquid film thickness between two approaching droplets using integrated optical interferometry and interface probe method. J. Colloid Interface Sci. 2025, 695, 137728. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  35. Zhang, J.; Shi, L.; Zhang, R.; Chen, J.; Wu, G. Spectral Interference Ellipsometry for Film Thickness Measurement on Transparent Substrate. Opt. Lasers Eng. 2023, 171, 107819. [Google Scholar] [CrossRef] [Scilit]
  36. Kim, S.; Kim, Y.; Shin, S.-C.; Hibino, K.; Sugita, N. Interferometric thickness measurement of glass plate by phase-shifting analysis using wavelength scanning with elimination of bias phase error. Opt. Rev. 2021, 28, 48–57. [Google Scholar] [CrossRef] [Scilit]
  37. Hao, F.; Zhang, Y.; Xu, Z.; Zhang, Y.; Li, G. A novel method for three-dimensional liquid surface topography measurement: Combining binocular vision with deep learning-based digital image correlation. Opt. Lasers Eng. 2025, 191, 108993. [Google Scholar] [CrossRef] [Scilit]
  38. Cen, Z.; Zeng, K.; Liu, B.; Huang, L.; He, Z.; Xiong, H.; Han, D.; Zeng, Y. Three-dimensional color camera based on optical coherence tomography. Opt. Commun. 2021, 478, 126363. [Google Scholar] [CrossRef] [Scilit]
  39. Ghim, Y.S.; Rhee, H.G.; Davies, A. Simultaneous measurements of top surface and its underlying film surfaces in multilayer film structure. Sci. Rep. 2017, 7, 11843. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  40. Miao, F.; Ahn, S.; Kim, Y. Precise Measurement of the Surface Shape of Silicon Wafer by Using a New Phase-Shifting Algorithm and Wavelength-Tuning Interferometer. Appl. Sci. 2020, 10, 3250. [Google Scholar] [CrossRef] [Scilit]
  41. Mo, J.; Chen, P.; Zhou, Y.; Peng, L.; Yu, Y. Adaptive Wavelength-Tuning Phase-Shifting Interferometry under Blind Cavity Length and Sampling Frequency. Opt. Commun. 2026, 603, 132776. [Google Scholar] [CrossRef] [Scilit]
  42. Nguyen, M.T.; Park, I.K.; Rhee, H.G.; Ghim, Y.S. Deep-learning based fringe-print-through error and noise removal for dynamic phase-shifting interferometry. Opt. Laser Technol. 2025, 191, 113313. [Google Scholar] [CrossRef] [Scilit]
  43. Gomez-Conde, J.C.; Meneses-Fabian, C. Real-time phase step measurement using the volume enclosed by a surface algorithm in self-calibrating phase-shifting interferometry. Measurement 2020, 153, 107412. [Google Scholar] [CrossRef] [Scilit]
  44. Takahashi, Y. Effect of phase error in phase-shifting interferometer. Appl. Mech. Mater. 2019, 888, 11–16. [Google Scholar] [CrossRef] [Scilit]
  45. Cheng, J.L.; Gao, Z.S.; Yuan, Q.; Wang, K.L.; Xu, L.P. Carrier squeezing interferometry with π/4 phase shift: Phase extraction in the presence of multi-beam interference. Appl. Opt. 2016, 55, 1920–1928. [Google Scholar] [CrossRef] [Scilit] [PubMed]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.