Dual-wavelength phase imaging method and device based on deep learning, equipment and medium

Abstract

Embodiments of the present application provide a dual-wavelength phase imaging method and device based on deep learning, equipment and medium. The method comprises: acquiring a single-channel normalized dual-wavelength interferogram; inputting the single-channel normalized dual-wavelength interferogram into a pre-trained deep learning network model for phase imaging, and outputting a single-channel normalized absolute phase image, wherein the deep learning network model adopts an R-Net network architecture, the deep learning network model comprises an encoder, a bridging bottleneck module, a decoder, a multi-scale fusion module, a refinement reconstruction module and an output end, and the down-sampling module of the encoder, the up-sampling module of the decoder, the multi-scale fusion module and the refinement reconstruction module all contain a residual double convolution module. Based on this, the absolute phase image finally output by the embodiments of the present application has higher continuity, integrity and reconstruction accuracy, and has stronger anti-noise ability and better phase recovery stability under complex fringes, especially under the condition of a noisy interferogram.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of phase retrieval technology, and in particular to a dual-wavelength phase imaging method, apparatus, device and medium based on deep learning. Background Technology

[0002] Quantitative phase imaging, with its advantages of being non-contact, non-destructive, high-precision, high-sensitivity, and capable of quantitative analysis, is particularly suitable for imaging weakly absorbing phase targets such as biological cells, and has therefore become an important direction in modern optical imaging. Around phase imaging, various technical approaches have been developed, including iterative algorithms, off-axis interferometry, digital holography, and white-light diffraction phase microscopy. The core of these methods is to first acquire an intensity map, and then obtain the phase information of the sample through phase retrieval. The accuracy and speed of phase retrieval directly determine the performance of the entire phase imaging system.

[0003] While traditional dual-wavelength interferometry can reduce the difficulties caused by phase wrapping and obtain continuous phase information, it usually still requires acquiring multiple frames of interferometric intensity maps first, then combining three or more phase shift calculations to separate the interference information of the two wavelengths, and finally performing phase unwrapping or continuous phase recovery. This results in a large number of acquisition frames, time-consuming measurement and calculation processes, and unstable phase recovery. In recent years, convolutional networks, represented by U-Net, have been widely used in the field of phase retrieval; however, traditional U-Net has the following significant drawbacks: 1. Limited feature extraction and feature fusion capabilities: Traditional U-Net mainly adopts the structure of "encoder-decoder + same-scale skip connection". Its convolution module is relatively standard, and the fusion method of deep semantic features and shallow detail features is relatively simple. It still lacks the ability to represent large-scale phase trends, small-scale local details and local fringe distortion that exist simultaneously in dual-wavelength interference fringes.

[0004] 2. Insufficient noise resistance and generalization ability: Dual-wavelength interferograms are usually affected by multiple factors such as Gaussian noise, speckle noise, uneven illumination, fringe contrast fluctuations, and slight defocus blur. Traditional U-Net is less adaptable to composite noise and complex fringe perturbations.

[0005] 3. Insufficient utilization of cross-scale information in dual-wavelength interferometry: Although the traditional U-Net has a basic multi-layer coding and decoding structure, it lacks depth interaction between features of different scales and is difficult to simultaneously take into account global phase trend judgment and local detail correction. Therefore, it is easy to have local misjudgment or insufficient continuity in absolute phase recovery. Summary of the Invention

[0006] This invention provides a deep learning-based dual-wavelength phase imaging method, apparatus, device, and medium. Residual double convolution modules are employed in each feature extraction unit to enhance feature reuse capability and deep gradient transfer stability. Simultaneously, at the decoding end, decoded features from different scales in the multi-scale fusion module and the thinning reconstruction module are uniformly mapped, upsampled, aligned, and fused to balance shallow local fringe details with deep global phase semantic information. At the output end, an initial absolute phase is first obtained through a coarse prediction head, and then the prediction result is compensated and corrected using a thinning output head. This results in a final output absolute phase map with higher continuity, completeness, and reconstruction accuracy, exhibiting stronger noise resistance and better phase recovery stability under complex fringe conditions, especially noisy interferograms.

[0007] In a first aspect, embodiments of the present invention provide a deep learning-based dual-wavelength phase imaging method, comprising: Obtain a single-channel normalized dual-wavelength interferogram; The single-channel normalized dual-wavelength interferogram is input into a pre-trained deep learning network model for phase imaging, outputting a single-channel normalized absolute phase map. The deep learning network model employs an R-Net architecture and includes an encoder, a bottleneck bridging module, a decoder, a multi-scale fusion module, a refinement and reconstruction module, and an output. The encoder includes an input residual convolution module and three cascaded downsampling modules. The decoder includes three cascaded upsampling modules. Both the fusion module and the refinement reconstruction module include a residual double convolution module, which includes a main branch and a shortcut branch. The outputs of the main branch and the shortcut branch are added element-wise and then output through an activation function to form a residual connection structure. The multi-scale fusion module and the refinement reconstruction module are connected sequentially between the end of the decoder and the output end. The output end includes a coarse prediction head and a refinement output head. The output end is used to generate the single-channel normalized absolute phase map based on the coarse prediction result output by the coarse prediction head and the refinement compensation result output by the refinement output head.

[0008] In some embodiments, the method for constructing the training dataset used by the deep learning network model during the training phase includes: On a two-dimensional Cartesian coordinate grid, the basic object shape is constructed using composite surface functions, and the basic phase distribution is generated. Define a first wavelength and a second wavelength, and calculate the phase term of the first wavelength and the phase term of the second wavelength based on the basic phase distribution; By introducing a phase shift, a dual-wavelength interferogram is generated based on a dual-wavelength interferometric imaging model; Calculate the difference between the phase term of the first wavelength and the phase term of the second wavelength to obtain the continuous phase; The continuous phases are globally normalized to generate absolute phase labels; The training dataset is constructed by using the dual-wavelength interferogram as the input sample and the absolute phase label as the output ground truth.

[0009] In some embodiments, the input residual convolution module is used to perform initial feature mapping on the single-channel normalized dual-wavelength interferogram; the downsampling module is composed of a max pooling layer and a residual double convolution module connected in series; the upsampling module sequentially includes a transposed convolutional layer with a stride of 2 and a kernel size of 2×2, a skip connection, and the residual double convolution module.

[0010] In some embodiments, the main branch sequentially includes a first 3×3 two-dimensional convolutional layer, a first set of normalization layers, a linear rectified activation function, a second 3×3 two-dimensional convolutional layer, and a second set of normalization layers; when the number of input channels is the same as the number of output channels, the shortcut branch uses an identity mapping; when the number of input channels is different from the number of output channels, the shortcut branch uses a convolutional layer with a kernel size of 1×1 to project the channel dimension, so that the output of the shortcut branch is consistent with the output of the main branch in the channel dimension.

[0011] In some embodiments, the processing steps of the multi-scale fusion module include: Multiple decoded feature maps are received from the decoder, respectively; The channel dimensions of multiple decoded feature maps are uniformly projected to a preset dimension through a 1×1 convolutional layer; The spatial resolution of the multiple decoded feature maps is upsampled to the highest resolution of the decoded feature map using bilinear interpolation. Figure 1 To; The multiple decoded feature maps after resolution alignment are concatenated along the channel dimension to obtain the concatenated feature map; The concatenated features are fused using the residual double convolution module to output multi-scale fused features.

[0012] In some embodiments, the processing steps at the output terminal include: The highest-resolution decoded feature map output by the decoder is input into the coarse prediction head, and a coarse prediction absolute phase map is generated through a 1×1 convolutional layer. The highest resolution decoded feature map is concatenated with the multi-scale fusion feature output by the multi-scale fusion module, and then input into the refinement and reconstruction module for feature refinement to obtain the refined feature. The refined features are input into the refined output head, and a refined compensation term is generated through a 1×1 convolutional layer. The coarse predicted absolute phase map is added element-wise to the refined compensation term to obtain the final output feature, and the output is restricted to the [0,1] interval by the Sigmoid activation function to obtain the single-channel normalized absolute phase map.

[0013] In some embodiments, the training method for the deep learning network model includes: The normalized phase mean square error, the wrap phase consistency error, and the gradient error are determined. The normalized phase mean square error is used to constrain the scalar regression error between the network output and the true normalized absolute phase. The wrap phase consistency error is used to constrain the predicted absolute phase and the true absolute phase to be consistent in the wrapping sense. The gradient error is used to constrain the gradient distribution of the predicted phase map and the true phase map to be consistent in the horizontal and vertical directions. A composite loss function is constructed based on the normalized phase mean square error, the wrapping phase consistency error, and the gradient error. The deep learning network model is trained based on the composite loss function to obtain the trained deep learning network model.

[0014] Secondly, embodiments of the present invention also provide a deep learning-based dual-wavelength phase imaging device, the device comprising: The acquisition module is used to acquire a single-channel normalized dual-wavelength interferogram; The processing module is used to input the single-channel normalized dual-wavelength interferogram into a pre-trained deep learning network model for phase imaging, and output a single-channel normalized absolute phase map. The deep learning network model adopts an R-Net network architecture and includes an encoder, a bottleneck bridging module, a decoder, a multi-scale fusion module, a thinning reconstruction module, and an output terminal. The encoder includes an input residual convolution module and three cascaded downsampling modules. The decoder includes three cascaded upsampling modules. Both the scale fusion module and the refinement reconstruction module include a residual double convolution module. The residual double convolution module includes a main branch and a shortcut branch. The outputs of the main branch and the shortcut branch are added element-wise and then output through an activation function to form a residual connection structure. The multi-scale fusion module and the refinement reconstruction module are connected sequentially between the end of the decoder and the output end. The output end includes a coarse prediction head and a refinement output head. The output end is used to generate the single-channel normalized absolute phase map based on the coarse prediction result output by the coarse prediction head and the refinement compensation result output by the refinement output head.

[0015] Thirdly, embodiments of the present invention also provide an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the deep learning-based dual-wavelength phase imaging method as described in the first aspect.

[0016] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions for performing the deep learning-based dual-wavelength phase imaging method as described in the first aspect.

[0017] The present invention provides a deep learning-based dual-wavelength phase imaging method, apparatus, device, and medium. The deep learning-based dual-wavelength phase imaging method includes: acquiring a single-channel normalized dual-wavelength interferogram; inputting the single-channel normalized dual-wavelength interferogram into a pre-trained deep learning network model for phase imaging; and outputting a single-channel normalized absolute phase map. The deep learning network model adopts an R-Net network architecture and includes an encoder, a bottleneck bridging module, a decoder, a multi-scale fusion module, a thinning reconstruction module, and an output terminal. The encoder includes an input residual convolution module and three cascaded modules. The decoder includes a downsampling module and three cascaded upsampling modules. Each of these modules—downsampling, upsampling, multi-scale fusion, and refinement reconstruction—contains a residual double convolution module. This module includes a main branch and a shortcut branch. The outputs of the main branch and the shortcut branch are added element-wise and then output through an activation function to form a residual connection structure. The multi-scale fusion module and the refinement reconstruction module are sequentially connected between the decoder's end and its output. The output includes a coarse prediction head and a refinement output head. The output is used to generate a single-channel normalized absolute phase map based on the coarse prediction result from the coarse prediction head and the refinement compensation result from the refinement output head. Based on this, the embodiments of the present invention employ residual double convolution modules in the downsampling module, upsampling module, multi-scale fusion module, and thinning reconstruction module to enhance feature reuse capability and deep gradient transfer stability. Simultaneously, at the decoding end, the decoded features of different scales from the multi-scale fusion module and the thinning reconstruction module are uniformly mapped, upsampled, aligned, and fused to balance shallow local fringe details with deep global phase semantic information. At the output end, an initial absolute phase is first obtained through a coarse prediction head, and then the prediction result is compensated and corrected using a thinning output head, resulting in a final output absolute phase map with higher continuity, completeness, and reconstruction accuracy. This provides stronger noise resistance and better phase recovery stability under complex fringe conditions, especially noisy interferograms. Attached Figure Description

[0018] Figure 1 This is a flowchart of a deep learning-based dual-wavelength phase imaging method provided in one embodiment of the present invention; Figure 2 This is a partial image showing the training dataset; Figure 3 This is a schematic diagram of the structure of an R-Net provided in one embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a residual double convolution module provided in one embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a multi-scale fusion module provided in one embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a refined reconstruction module provided in one embodiment of the present invention; Figure 7 (a) is the noise-free interferogram, (b) is the corresponding absolute phase diagram, (c) is the predicted absolute phase diagram output by U-Net, (d) is the U-Net model error diagram, (e) is the predicted absolute phase diagram output by R-Net, and (f) is the R-Net error diagram. Figure 8 (a) is the interferogram of 15dB Gaussian white noise, (b) is the corresponding absolute phase diagram, (c) is the predicted absolute phase diagram of U-Net output, (d) is the U-Net model error diagram, (e) is the predicted absolute phase diagram of R-Net output, and (f) is the R-Net error diagram. Figure 9 This is a schematic diagram of a deep learning-based dual-wavelength phase imaging device provided in one embodiment of the present invention; Figure 10 This is a schematic diagram of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the following drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0021] In this embodiment of the invention, the terms "furthermore," "exemplarily," or "optionally" are used to indicate examples, illustrations, or explanations, and should not be construed as being more preferred or advantageous than other embodiments or designs. The use of terms such as "furthermore," "exemplarily," or "optionally" is intended to present the relevant concepts in a specific manner.

[0022] First, let's analyze some of the terms used in this invention: Phase recovery: In applications such as optical imaging, interferometry, and digital holography, detectors typically only record the intensity information of the light field, but cannot directly obtain its phase information. Phase recovery aims to reconstruct the true phase distribution of the light field under test based on the measured intensity distribution, combined with propagation models, interference constraints, or prior conditions, using appropriate mathematical methods. This allows for the further acquisition of the target's morphology, position, refractive index distribution, or other physical information.

[0023] Deep learning is a type of machine learning method that uses multi-layered neural networks as its core. It can automatically learn complex nonlinear mapping relationships between inputs and outputs through training on large amounts of data. Through multi-layer feature extraction and progressive representation learning, it can automatically mine low-level detailed features and high-level semantic features from raw data, thereby completing tasks such as classification, recognition, segmentation, reconstruction, and regression. U-Net is a typical deep learning convolutional neural network structure, commonly used for tasks such as image segmentation, image reconstruction, and phase retrieval. Its network typically consists of an encoder and a decoder: the encoder is responsible for extracting multi-scale features from the input data layer by layer, while the decoder is responsible for progressively restoring the spatial resolution and fusing shallow detailed information with deep semantic information through skip connections, thus achieving high-precision prediction of the target result. In phase retrieval or height reconstruction tasks, U-Net can learn the mapping relationship between interferograms, holograms, or complex field features and the target phase map and height map, representing a typical application of deep learning in the field of optical imaging.

[0024] To facilitate a more convenient description of the working principle of the embodiments of the present invention, the following introduction of relevant technical scenarios is given first.

[0025] Quantitative phase imaging, with its advantages of being non-contact, non-destructive, high-precision, high-sensitivity, and capable of quantitative analysis, is particularly suitable for imaging weakly absorbing phase targets such as biological cells, and has therefore become an important direction in modern optical imaging. Around phase imaging, various technical approaches have been developed, including iterative algorithms, off-axis interferometry, digital holography, and white-light diffraction phase microscopy. The core of these methods is to first acquire an intensity map, and then obtain the phase information of the sample through phase retrieval. The accuracy and speed of phase retrieval directly determine the performance of the entire phase imaging system.

[0026] While traditional dual-wavelength interferometry can reduce the difficulties caused by phase wrapping and obtain continuous phase information, it usually still requires acquiring multiple frames of interferometric intensity maps first, then combining three or more phase shift calculations to separate the interference information of the two wavelengths, and finally performing phase unwrapping or continuous phase recovery. This results in a large number of acquisition frames, time-consuming measurement and calculation processes, and unstable phase recovery. In recent years, convolutional networks, represented by U-Net, have been widely used in the field of phase retrieval; however, traditional U-Net has the following significant drawbacks: 1. Limited feature extraction and feature fusion capabilities: Traditional U-Net mainly adopts the structure of "encoder-decoder + same-scale skip connection". Its convolution module is relatively standard, and the fusion method of deep semantic features and shallow detail features is relatively simple. It still lacks the ability to represent large-scale phase trends, small-scale local details and local fringe distortion that exist simultaneously in dual-wavelength interference fringes.

[0027] 2. Insufficient noise resistance and generalization ability: Dual-wavelength interferograms are usually affected by multiple factors such as Gaussian noise, speckle noise, uneven illumination, fringe contrast fluctuations, and slight defocus blur. Traditional U-Net is less adaptable to composite noise and complex fringe perturbations.

[0028] 3. Insufficient utilization of cross-scale information in dual-wavelength interferometry: Although the traditional U-Net has a basic multi-layer coding and decoding structure, it lacks depth interaction between features of different scales and is difficult to simultaneously take into account global phase trend judgment and local detail correction. Therefore, it is easy to have local misjudgment or insufficient continuity in absolute phase recovery.

[0029] Based on this, the present invention provides a deep learning-based dual-wavelength phase imaging method, apparatus, device, and medium. The deep learning-based dual-wavelength phase imaging method includes: acquiring a single-channel normalized dual-wavelength interferogram; inputting the single-channel normalized dual-wavelength interferogram into a pre-trained deep learning network model for phase imaging, and outputting a single-channel normalized absolute phase map. The deep learning network model adopts an R-Net network architecture and includes an encoder, a bottleneck bridging module, a decoder, a multi-scale fusion module, a thinning reconstruction module, and an output terminal. The encoder includes an input residual convolution module and three cascaded downsampling modules, and the decoder includes three cascaded... The upsampling module, downsampling module, multi-scale fusion module, and refinement reconstruction module all include a residual double convolution module. The residual double convolution module includes a main branch and a shortcut branch. The outputs of the main branch and the shortcut branch are added element-wise and then output through an activation function to form a residual connection structure. The multi-scale fusion module and the refinement reconstruction module are connected sequentially between the end of the decoder and the output end. The output end includes a coarse prediction head and a refinement output head. The output end is used to generate a single-channel normalized absolute phase map based on the coarse prediction result output by the coarse prediction head and the refinement compensation result output by the refinement output head. Based on this, the embodiments of the present invention employ residual double convolution modules in the downsampling module, upsampling module, multi-scale fusion module, and thinning reconstruction module to enhance feature reuse capability and deep gradient transfer stability. Simultaneously, at the decoding end, the decoded features of different scales from the multi-scale fusion module and the thinning reconstruction module are uniformly mapped, upsampled, aligned, and fused to balance shallow local fringe details with deep global phase semantic information. At the output end, an initial absolute phase is first obtained through a coarse prediction head, and then the prediction result is compensated and corrected using a thinning output head, resulting in a final output absolute phase map with higher continuity, completeness, and reconstruction accuracy. This provides stronger noise resistance and better phase recovery stability under complex fringe conditions, especially noisy interferograms.

[0030] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0031] like Figure 1 As shown, Figure 1 This is a flowchart of a deep learning-based dual-wavelength phase imaging method provided in an embodiment of the present invention. The deep learning-based dual-wavelength phase imaging method may include, but is not limited to, steps S101 to S102.

[0032] Step S101: Obtain a single-channel normalized dual-wavelength interferogram; Step S102: Input the single-channel normalized dual-wavelength interferogram into a pre-trained deep learning network model for phase imaging, and output a single-channel normalized absolute phase map. The deep learning network model adopts the R-Net network architecture and includes an encoder, a bridging bottleneck module, a decoder, a multi-scale fusion module, a refinement and reconstruction module, and an output. The encoder includes an input residual convolution module and three cascaded downsampling modules. The decoder includes three cascaded upsampling modules. The downsampling module, upsampling module, multi-scale fusion module, and refinement and reconstruction module all contain residual double convolution modules. The residual double convolution module includes a main branch and a shortcut branch. The output of the main branch and the output of the shortcut branch are added element-wise and output through an activation function to form a residual connection structure. The multi-scale fusion module and the refinement and reconstruction module are connected sequentially between the decoder's end and the output. The output includes a coarse prediction head and a refinement output head. The output is used to generate a single-channel normalized absolute phase map based on the coarse prediction result output by the coarse prediction head and the refinement compensation result output by the refinement output head.

[0033] Understandably, in order to address the shortcomings of the standard U-Net, this invention proposes to adopt the R-Net network architecture for deep learning network models. The R-Net network architecture performs dual-wavelength interferometric absolute phase recovery on the basis of the standard U-Net to improve feature extraction capability, multi-scale fusion capability and final phase reconstruction accuracy.

[0034] It is understood that the deep learning network model of this invention adopts the R-Net network. The R-Net network includes an encoder, a bottleneck bridging module, a decoder, a multi-scale fusion module, a refinement and reconstruction module, and an output terminal. The encoder includes an input residual convolution module and three cascaded downsampling modules. The decoder includes three cascaded upsampling modules. The downsampling module, upsampling module, multi-scale fusion module, and refinement and reconstruction module all include residual double convolution modules. The residual double convolution module includes a main branch and a shortcut branch. The output of the main branch and the output of the shortcut branch are added element-wise and output through an activation function to form a residual connection structure. The multi-scale fusion module and the refinement and reconstruction module are connected sequentially between the decoder's end and the output terminal. The output terminal includes a coarse prediction head and a refinement output head. The output terminal is used to generate a single-channel normalized absolute phase map based on the coarse prediction result output by the coarse prediction head and the refinement compensation result output by the refinement output head. Based on this, inputting the single-channel normalized dual-wavelength interferogram into the pre-trained deep learning network model for accurate and fast phase recovery can predict the absolute phase map from the dual-wavelength interferogram in one step.

[0035] Understandably, during the training phase of deep learning network models, it is necessary to construct a training dataset to train the R-Net network. The deep learning-based dual-wavelength interferometric absolute phase retrieval method requires a large number of paired samples of "input interferogram - output absolute phase map" for supervised training. However, in actual measurements, it is difficult to obtain a large number of rigorously registered dual-wavelength interferograms and their corresponding true absolute phase values. Therefore, this invention uses a combination of physical modeling and numerical simulation to construct the training dataset. A portion of the constructed training dataset is shown below. Figure 2 As shown.

[0036] This invention focuses solely on constructing datasets for complex fringe patterns. First, a fixed-size Cartesian coordinate grid, preferably 512×512, is established in two-dimensional space. Then, a composite surface similar to a Peaks function is used to construct the basic object's shape, simulating the spatial phase distribution corresponding to complex surface undulations. Since Peaks-like surfaces simultaneously contain multiple local peaks and valleys, nonlinear variation regions, and complex spatial topologies, they can more effectively simulate the phase characteristics of complex fringe scenes in dual-wavelength interferometric phase retrieval.

[0037] Let the composite surface be P(x,y), and then introduce a random modulation coefficient nf, the preferred range of which is:

[0038] The phase of the fundamental object can then be expressed as:

[0039] In the dual-wavelength interferometric imaging model, let the first wavelength and the second wavelength be:

[0040] The phase terms corresponding to the two wavelengths are as follows:

[0041]

[0042] Furthermore, to simulate dual-wavelength phase-shift interference, this invention introduces phase shift amounts corresponding to two wavelengths. and In the preferred embodiment, a common-path discrete five-step phase-shift mode is employed. That is, firstly, based on the dual-wavelength synthesized wavelength...

[0043] Generate discrete path step size Then calculate the phase shift corresponding to the two wavelengths:

[0044]

[0045] Based on the above physical model, a two-wavelength interferogram is constructed:

[0046] in, For background items, and These represent the fringe modulation terms corresponding to the two wavelengths. Preferably, both the background term and the modulation term incorporate Gaussian envelope modulation to more closely approximate the spatial intensity distribution in actual dual-wavelength interferometric imaging.

[0047] Subsequently, the above-mentioned theoretical continuous phase is defined as:

[0048] To enable learning from different samples at a uniform scale, this invention preferably employs global continuous phase range normalization, and further maps it to... The interval is used to obtain the absolute phase label:

[0049] Then normalize it to:

[0050] in, and Let represent the minimum and maximum values ​​of the global continuous phase in the complex striped dataset, respectively. Thus, the network's learning objective is not to wrap the phase, but rather to learn the normalized continuous absolute phase.

[0051] Understandably, in related technologies, the U-Net network is characterized by a symmetrical setup of the decoder and encoder, containing a total of four upsampling recovery layers. Each decoding layer first performs upsampling through a 2×2 transposed convolution, then performs a skip connection with the feature map of the corresponding scale from the encoder. Next, it passes through a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function to complete feature fusion and spatial detail restoration. Finally, a 1×1 convolution compresses the feature channels to 1, outputting a normalized absolute phase prediction map. Its basic loss function is the mean squared error loss of the normalized absolute phase.

[0052] in, The normalized absolute phase predicted by the network. For true normalized absolute phase, N is the total number of pixels.

[0053] like Figure 3As shown, the R-Net network proposed in this invention uses an encoder and a decoder as its basic structure, and takes a single-channel normalized dual-wavelength interferogram as input and a single-channel normalized absolute phase map as output. However, it differs from the standard U-Net. Compared to other methods, it has improved in terms of basic feature extraction units, encoding and decoding feature transfer methods, multi-scale fusion methods, and final output reconstruction methods, thereby improving the feature extraction capability, multi-scale context information integration capability, and final absolute phase recovery accuracy under complex dual-wavelength interference fringe conditions. The R-Net network preferably includes an input residual convolution module inc, three downsampling modules down1, down2, and down3, a bridging bottleneck module bridge, three upsampling modules up1, up2, and up3, a multi-scale fusion module MultiScaleFusion, a refinement reconstruction module refine, a coarse prediction head coarse_head, and a refinement output head outc. In the preferred embodiment, the number of basic channels of the network is set to 64, so that the feature channels in the network backbone change sequentially from 64 to 128 to 256 to 512, so as to gradually increase the feature representation dimension while expanding the receptive field step by step.

[0054] The input single-channel normalized dual-wavelength interferogram first enters the input residual convolution module (inc) for initial feature mapping. Then, it sequentially passes through three cascaded downsampling modules (down1, down2, and down3) for multi-scale hierarchical feature extraction to obtain semantic feature representations at different spatial resolutions. Each downsampling module consists of a max-pooling layer connected in series with a residual double convolution module. Specifically, the max-pooling layer first compresses the spatial resolution of the input feature map, and then the residual double convolution module extracts deep features from the downsampled feature map. This process compresses the spatial size of the feature map while progressively enhancing the network's ability to express global structural information, local phase change trends, and complex texture regions in complex dual-wavelength interferogram fringes. This invention uniformly uses the residual double convolution module (ResidualDoubleConv) as the basic feature extraction unit in both the encoder and decoder. Figure 4As shown, the main branch of the ResidualDoubleConv module preferably includes, in sequence, a 2D convolutional layer with a kernel size of 3×3, a group normalization (GN) layer, a linear rectified activation function (ReLU) layer, another 2D convolutional layer with a kernel size of 3×3, and another group normalization layer. A shortcut branch is set outside the main branch. When the number of input channels is the same as the number of output channels, the shortcut branch preferably uses an identity mapping. When the number of input channels is different from the number of output channels, the shortcut branch preferably uses a convolutional layer with a kernel size of 1×1 to project the channel dimension, so that the output of the shortcut branch is consistent with the output of the main branch in the channel dimension. Finally, the output of the main branch and the output of the shortcut branch are added element-wise and output through an activation function to form a residual connection structure. This enhances the reuse capability of shallow detail information, improves the gradient propagation stability in deep networks, and enhances the accuracy and robustness of phase detail recovery in complex interference fringe regions.

[0055] This invention establishes a bridging bottleneck module, which preferably consists of two consecutively cascaded residual double convolutional modules. This module is used to further refine deep semantic features at the lowest spatial resolution and highest feature abstraction level, thereby enhancing the network's comprehensive modeling ability for global phase distribution patterns, nonlinear topological structures, and complex background disturbances in complex dual-wavelength interference fringes. Furthermore, in some preferred embodiments, a lightweight channel attention mechanism can be further introduced within the residual double convolutional module to adaptively enhance key phase feature channels and suppress irrelevant background channel responses.

[0056] In the decoding stage, the R-Net network's decoder consists of three symmetrically distributed upsampling modules up1, up2, and up3. Each upsampling module sequentially includes a transposed convolutional layer (ConvTranspose2d) with a stride of 2 and a kernel size of 2×2, a skip connection, and a residual double convolutional module. Low-resolution deep features are first upsampled through transposed convolution to restore spatial resolution, and then concatenated with the feature map of the corresponding scale in the encoder along the channel dimension to achieve the fusion of shallow high-resolution detail information and deep low-resolution semantic information. Afterward, the residual double convolutional module reconstructs and refines the concatenated features, thereby restoring the spatial topology of the feature map step by step and outputting the decoded feature maps u1, u2, and u3 at the corresponding scale. Through this structural design, the local edges, subtle textures, and global phase continuity of complex stripe regions can be taken into account during the phase reconstruction process.

[0057] This invention introduces an additional multi-scale fusion module, MultiScaleFusion, at the decoding end, such as... Figure 5 As shown, the multi-scale fusion module simultaneously receives feature maps u1, u2, and u3 from different decoding layers, and projects them onto the channels through convolutional layers with a kernel size of 1×1 to map features of different scales to the same feature channel dimension. Then, bilinear interpolation is used to upsample the lower resolution features to the same spatial resolution as the highest resolution decoded features, and the features are stitched together in the channel dimension. Finally, the residual double convolution module is used to fuse the stitched features, thereby constructing a multi-scale fusion feature representation that takes into account shallow local texture details, mid-level structural features, and deep global semantic information. This enhances the network's ability to jointly perceive and comprehensively express phase features of different spatial scales in complex dual-wavelength interference fringes during the final reconstruction stage.

[0058] After obtaining the above-mentioned multi-scale fusion features, the present invention further sets up a refinement reconstruction module, such as... Figure 6 As shown, its input is the channel concatenation result of the highest resolution decoded feature u3 and the output feature of the multi-scale fusion module. The refine module preferably uses a residual double convolutional structure to further refine the concatenated feature to generate a high-quality feature representation for refinement compensation. At the output end, the present invention adopts a dual-head output structure of "coarse prediction + refinement correction". That is, firstly, the coarse prediction head coarse_head uses a convolutional layer with a kernel size of 1×1 to perform initial absolute phase prediction on the highest resolution decoded feature u3 to recover the overall trend information of the phase field. Then, the refinement output head outc uses a convolutional layer with a kernel size of 1×1 to refine the feature. The module outputs a refined feature to generate a compensation term to correct local peak and valley regions, complex fringe boundary regions, and high-frequency detail regions. Finally, the coarse prediction result and the refined compensation result are added element-wise to obtain the final output logits. The output is then restricted to the [0,1] interval by the Sigmoid activation function to output the final normalized absolute phase map. This structure allows the network to first recover the global distribution of the absolute phase and then further correct the detail errors in complex local regions, making it more suitable for complex dual-wavelength interferometric absolute phase recovery tasks.

[0059] It is understandable that, regarding the training objective, this invention preferably employs a composite loss function consisting of normalized phase error, wrapping phase consistency error, and gradient error to optimize the training of the R-Net network, the expression of which is:

[0060] Wherein, the normalized phase mean square error is:

[0061] Used to constrain the scalar regression error between the network output and the true normalized absolute phase; The package phase consistency loss is as follows:

[0062] This is used to constrain the predicted absolute phase to remain consistent with the true absolute phase in a wrapper sense; where...

[0063] Principal values ​​representing phase difference are wrapped in a computational operation. The gradient loss is:

[0064] To constrain the consistency of gradient distribution between the predicted phase map and the true phase map in the horizontal and vertical directions, thereby improving the consistency of the final recovered phase in terms of spatial continuity, edge details, and local structure, in a preferred embodiment, the loss weight can be set as follows:

[0065] This allows the network to maintain the accuracy of absolute phase numerical regression while further considering the consistency of the wrapped phase and the consistency of the spatial gradient structure, making it more suitable for absolute phase recovery tasks under complex dual-wavelength interference fringe conditions.

[0066] To evaluate the performance of the model on multiple scales, all test results in this invention are comprehensively evaluated using RMSE, PSNR, and SSIM metrics.

[0067] Figure 7 and Figure 8 In the diagram, (a) represents the noise-free interferogram, i.e., the input of the model; (b) represents the absolute phase diagram corresponding to (a); (c) and (d) represent the output and error diagram of U-Net, respectively; and (e) and (f) represent the output and error diagram of R-Net, respectively.

[0068] As shown in Tables 1 and 2 below, the proposed model R-Net outperforms U-Net in all evaluation metrics. The experimental results fully demonstrate the advantages of the proposed network in terms of noise resistance, generalization, and robustness.

[0069] Table 1. Average PSNR (dB), RMSE, and SSIM of Phase Retrieval in Noise-Free Interferogram Model

[0070] Table 215 dB Gaussian white noise interferogram model phase retrieval average PSNR (dB), RMSE, SSIM

[0071] Understandably, the purpose of this invention is to address the shortcomings of the standard U-Net by providing a phase retrieval method that can directly establish a nonlinear mapping relationship between a dual-wavelength interferogram and an absolute phase map. This reduces the problem of error accumulation during traditional multi-step phase extraction and unwrapping processes. The network structure possesses stronger multi-scale feature extraction and fusion capabilities, improving the joint representation of the global phase trend, local detail texture, and complex boundary regions of dual-wavelength interferogram fringes, thereby enhancing the accuracy of absolute phase retrieval. This dual-wavelength interferometric phase retrieval method is also more robust to combined interferences such as noise, uneven illumination, fringe contrast variations, and local blurring, enhancing the network's generalization ability between simulated and real measurement data. It is a deep learning method suitable for high-quality, fast absolute phase retrieval of dual-wavelength interferograms, meeting the accuracy, efficiency, and robustness requirements of applications such as precision optical inspection, 3D topography measurement, and phase imaging.

[0072] It is understandable that, compared with the standard U-Net, the R-Net proposed in this invention does not simply employ a simple encoder-decoder and symmetric skip connection structure, but further introduces an output mechanism that combines residual feature extraction, multi-scale decoding fusion, and coarse prediction with refinement correction in the phase retrieval task. Specifically, this invention employs a residual double convolution structure in each feature extraction unit to enhance feature reuse capability and deep gradient propagation stability; simultaneously, at the decoding end, it performs unified mapping, upsampling alignment, and fusion of decoded features at different scales, thereby taking into account both shallow local fringe details and deep global phase semantic information; at the output end, the initial absolute phase is first obtained through the coarse prediction head, and then the prediction result is compensated and corrected using the refinement branch, so that the final output absolute phase map has higher continuity, integrity, and reconstruction accuracy.

[0073] The R-Net of this invention is also more suitable for absolute phase reconstruction tasks in dual-wavelength interferometry under complex interference fringes and noisy conditions. On the one hand, the residual structure and optional lightweight attention mechanism help the network extract effective fringe region features more stably and suppress redundant background interference; on the other hand, by combining noise-enhanced training, the robustness of the network under low signal-to-noise ratio interferograms can be significantly improved. Experimental results show that, compared with the standard U-Net, this invention has stronger noise resistance and better phase reconstruction stability under complex fringes, especially noisy interferograms, and is therefore more suitable for absolute phase reconstruction in practical dual-wavelength interferometry scenarios.

[0074] In addition, such as Figure 9 As shown, one embodiment of the present invention also discloses a deep learning-based dual-wavelength phase imaging device, the device comprising: The acquisition module 110 is used to acquire a single-channel normalized dual-wavelength interferogram; The processing module 120 is used to input a single-channel normalized dual-wavelength interferogram into a pre-trained deep learning network model for phase imaging and output a single-channel normalized absolute phase map. The deep learning network model adopts the R-Net network architecture and includes an encoder, a bottleneck bridging module, a decoder, a multi-scale fusion module, a refinement and reconstruction module, and an output terminal. The encoder includes an input residual convolution module and three cascaded downsampling modules. The decoder includes three cascaded upsampling modules. The downsampling module, upsampling module, multi-scale fusion module, and refinement and reconstruction module all contain residual double convolution modules. The residual double convolution module includes a main branch and a shortcut branch. The output of the main branch and the output of the shortcut branch are added element-wise and output through an activation function to form a residual connection structure. The multi-scale fusion module and the refinement and reconstruction module are connected sequentially between the end of the decoder and the output terminal. The output terminal includes a coarse prediction head and a refinement output head. The output terminal is used to generate a single-channel normalized absolute phase map based on the coarse prediction result output by the coarse prediction head and the refinement compensation result output by the refinement output head.

[0075] The deep learning-based dual-wavelength phase imaging device of this invention is used to execute the deep learning-based dual-wavelength phase imaging method in the above embodiments. Its specific processing procedure is the same as that of the deep learning-based dual-wavelength phase imaging method in the above embodiments, and will not be described in detail here.

[0076] In addition, such as Figure 10 As shown, one embodiment of the present invention also discloses an electronic device, including: at least one processor 210; at least one memory 220 for storing at least one program; when the at least one program is executed by the at least one processor 210, it implements the deep learning-based dual-wavelength phase imaging method as in any of the preceding embodiments.

[0077] In addition, one embodiment of the present invention discloses a computer-readable storage medium storing computer-executable instructions for performing the deep learning-based dual-wavelength phase imaging method as described in any of the preceding embodiments.

[0078] The system architecture and application scenarios described in the embodiments of this invention are for the purpose of more clearly illustrating the technical solutions of the embodiments of this invention, and do not constitute a limitation on the technical solutions provided by the embodiments of this invention. As those skilled in the art will know, with the evolution of system architecture and the emergence of new application scenarios, the technical solutions provided by the embodiments of this invention are also applicable to similar technical problems.

[0079] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0080] In hardware implementations, the division between functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0081] The terms “component,” “module,” “system,” etc., used in this specification are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, or a computer. As illustrated, applications running on computing devices and computing devices can both be components. One or more components may reside in a process or execution thread, and components may be located on a single computer or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate, for example, via local or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, or a network, such as the Internet interacting with other systems via signals).

Claims

1. A deep learning-based dual-wavelength phase imaging method, comprising: Obtain a single-channel normalized dual-wavelength interferogram; The single-channel normalized dual-wavelength interferogram is input into a pre-trained deep learning network model for phase imaging, outputting a single-channel normalized absolute phase map. The deep learning network model employs an R-Net architecture and includes an encoder, a bottleneck bridging module, a decoder, a multi-scale fusion module, a refinement and reconstruction module, and an output. The encoder includes an input residual convolution module and three cascaded downsampling modules. The decoder includes three cascaded upsampling modules. Both the fusion module and the refinement reconstruction module include a residual double convolution module, which includes a main branch and a shortcut branch. The outputs of the main branch and the shortcut branch are added element-wise and then output through an activation function to form a residual connection structure. The multi-scale fusion module and the refinement reconstruction module are connected sequentially between the end of the decoder and the output end. The output end includes a coarse prediction head and a refinement output head. The output end is used to generate the single-channel normalized absolute phase map based on the coarse prediction result output by the coarse prediction head and the refinement compensation result output by the refinement output head.

2. The method according to claim 1, characterized in that, The methods used to construct the training dataset during the training phase of the deep learning network model include: On a two-dimensional Cartesian coordinate grid, the basic object shape is constructed using composite surface functions, and the basic phase distribution is generated. Define a first wavelength and a second wavelength, and calculate the phase term of the first wavelength and the phase term of the second wavelength based on the basic phase distribution; By introducing a phase shift, a dual-wavelength interferogram is generated based on a dual-wavelength interferometric imaging model; Calculate the difference between the phase term of the first wavelength and the phase term of the second wavelength to obtain the continuous phase; The continuous phases are globally normalized to generate absolute phase labels; The training dataset is constructed by using the dual-wavelength interferogram as the input sample and the absolute phase label as the output ground truth.

3. The method according to claim 1, characterized in that, The input residual convolution module is used to perform initial feature mapping on the single-channel normalized dual-wavelength interferogram; the downsampling module is composed of a max pooling layer and a residual double convolution module connected in series; the upsampling module sequentially includes a transposed convolutional layer with a stride of 2 and a kernel size of 2×2, a skip connection, and the residual double convolution module.

4. The method according to claim 1, characterized in that, The main branch sequentially includes a first 3×3 two-dimensional convolutional layer, a first set of normalization layers, a linear rectified activation function, a second 3×3 two-dimensional convolutional layer, and a second set of normalization layers. When the number of input channels is the same as the number of output channels, the shortcut branch uses an identity mapping. When the number of input channels is different from the number of output channels, the shortcut branch uses a convolutional layer with a kernel size of 1×1 to project the channel dimension, so that the output of the shortcut branch is consistent with the output of the main branch in the channel dimension.

5. The method according to claim 1, characterized in that, The processing steps of the multi-scale fusion module include: Multiple decoded feature maps are received from the decoder, respectively; The channel dimensions of multiple decoded feature maps are uniformly projected to a preset dimension through a 1×1 convolutional layer; The spatial resolution of the multiple decoded feature maps is upsampled to match that of the decoded feature map with the highest resolution by bilinear interpolation. The multiple decoded feature maps after resolution alignment are concatenated along the channel dimension to obtain the concatenated feature map. The concatenated features are fused using the residual double convolution module to output multi-scale fused features.

6. The method according to claim 1, characterized in that, The processing steps at the output end include: The highest-resolution decoded feature map output by the decoder is input into the coarse prediction head, and a coarse prediction absolute phase map is generated through a 1×1 convolutional layer. The highest resolution decoded feature map is concatenated with the multi-scale fusion feature output by the multi-scale fusion module, and then input into the refinement and reconstruction module for feature refinement to obtain the refined feature. The refined features are input into the refined output head, and a refined compensation term is generated through a 1×1 convolutional layer. The coarse predicted absolute phase map is added element-wise to the refined compensation term to obtain the final output feature, and the output is restricted to the [0,1] interval by the Sigmoid activation function to obtain the single-channel normalized absolute phase map.

7. The method according to claim 1, characterized in that, The training methods for the deep learning network model include: The normalized phase mean square error, the wrap phase consistency error, and the gradient error are determined. The normalized phase mean square error is used to constrain the scalar regression error between the network output and the true normalized absolute phase. The wrap phase consistency error is used to constrain the predicted absolute phase and the true absolute phase to be consistent in the wrapping sense. The gradient error is used to constrain the gradient distribution of the predicted phase map and the true phase map to be consistent in the horizontal and vertical directions. A composite loss function is constructed based on the normalized phase mean square error, the wrapping phase consistency error, and the gradient error. The deep learning network model is trained based on the composite loss function to obtain the trained deep learning network model.

8. A deep learning-based dual-wavelength phase imaging device, characterized in that, The device includes: The acquisition module is used to acquire a single-channel normalized dual-wavelength interferogram; The processing module is used to input the single-channel normalized dual-wavelength interferogram into a pre-trained deep learning network model for phase imaging, and output a single-channel normalized absolute phase map. The deep learning network model adopts an R-Net network architecture and includes an encoder, a bottleneck bridging module, a decoder, a multi-scale fusion module, a thinning reconstruction module, and an output terminal. The encoder includes an input residual convolution module and three cascaded downsampling modules. The decoder includes three cascaded upsampling modules. Both the scale fusion module and the refinement reconstruction module include a residual double convolution module. The residual double convolution module includes a main branch and a shortcut branch. The outputs of the main branch and the shortcut branch are added element-wise and then output through an activation function to form a residual connection structure. The multi-scale fusion module and the refinement reconstruction module are connected sequentially between the end of the decoder and the output end. The output end includes a coarse prediction head and a refinement output head. The output end is used to generate the single-channel normalized absolute phase map based on the coarse prediction result output by the coarse prediction head and the refinement compensation result output by the refinement output head.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the deep learning-based dual-wavelength phase imaging method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions for performing the deep learning-based dual-wavelength phase imaging method as described in any one of claims 1 to 7.

Images