Axial multilayer phase information reconstruction method based on complex valued neural network
Through a method based on complex-valued neural networks, the speckle complex amplitude field is used as prior information to construct a multi-layer phase information reconstruction network of complex domain calculation units, which solves the problem of insufficient modeling ability of traditional methods in axial multi-layer phase reconstruction and achieves a reconstruction effect with higher accuracy and robustness.
Patent Information
- Application Number
- CN202511097873.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-06
- Publication Date
- 2025-10-17
AI Technical Summary
When existing scattering medium imaging technology reconstructs axial multi-layer phase information, traditional methods have insufficient ability to model complex signals and find it difficult to effectively utilize the complex characteristics of the light field, resulting in unsatisfactory reconstruction results.
A method based on complex-valued neural networks is adopted. By constructing a multi-layer phase information reconstruction network, using the speckle complex amplitude field as prior information, introducing a complex domain calculation unit, performing end-to-end feature mapping learning, and optimizing the network parameters to reconstruct axial multi-layer phase information.
It significantly improves the reconstruction accuracy and robustness of multi-layer phase information, simplifies the reconstruction process, and has good anti-interference ability and scalability.
Smart Images
Figure CN120807690A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of scattering medium imaging, and more particularly relates to an axial multi-layer phase information reconstruction method based on a complex-valued neural network. BACKGROUND
[0002] Scattering medium imaging is an important imaging technology that uses internal scattering information of a medium to obtain the structure and characteristics of a target. In this process, electromagnetic waves, sound waves or light waves interact with inhomogeneous structures in the medium during propagation, resulting in scattering phenomena. The generated scattering signals contain information about the target, which can be used to reconstruct the structure and physical properties of the target. Currently, scattering medium imaging technology has been widely used in geological exploration, medical imaging, material detection, environmental monitoring and security inspection, and other fields. In terms of technical path, existing scattering medium imaging mainly includes two categories: wavefront shaping and optical effect-based methods. The former mainly relies on optical phase conjugation technology and iterative algorithms such as feedback optimization to control the incident wavefront, achieving effective imaging in a scattering environment; the latter usually analyzes the autocorrelation characteristics of speckle patterns to extract the Fourier amplitude information of the target, and combines phase recovery algorithms to complete target reconstruction. However, traditional optical reconstruction methods have strong dependence on experimental conditions and strict preconditions, and are often limited by environmental changes and system complexity in practical applications, with deficiencies in imaging efficiency and robustness. In recent years, with the development of deep learning, researchers have begun to introduce it into scattering medium imaging by constructing end-to-end feature modeling networks to replace explicit physical modeling of medium propagation. This type of method directly learns the complex mapping relationship between the input and output, and has shown great potential in improving reconstruction efficiency and accuracy. Among them, convolutional neural networks (CNN) are widely used in reconstruction tasks from scattering images to original light field information, and have achieved remarkable results. Although CNN-based methods have made progress in scattering imaging, they still face several challenges. First, most traditional network structures are designed for light field intensity information, and do not effectively utilize the complex characteristics of light field, ignoring the potential correlation between amplitude and phase, which can easily cause loss of key information. Second, for multi-layer information with axial depth structure, real number domain networks are difficult to model the complex depth correlation, resulting in unsatisfactory reconstruction results. How to fully express the structural characteristics of complex amplitude in the network and accurately analyze the multi-layer depth distribution information is a core problem and key challenge in current high-precision scattering imaging research. SUMMARY
[0003] In view of the defects of the prior art, the application provides an axial multilayer phase information reconstruction method based on a complex-valued neural network, aiming at realizing reconstruction of multilayer phase information when the multilayer phase information with axial depth distribution is transmitted in a complex medium, breaking through the limitation of poor modeling capability of a complex signal by a traditional reconstruction method, and filling the gap in the related art.
[0004] The technical scheme for solving the above technical problems of the application is:
[0005] An axial multilayer phase information reconstruction method based on a complex-valued neural network comprises the following steps
[0006] S1, a MATLAB simulation software is used to simulate and calculate a complex amplitude field obtained by light passing through multilayer phase information placed at different axial relative distances;
[0007] S2, according to the complex amplitude field obtained in step S1, superpixel coding is used for holographic modulation, and then transmission is completed via a scattering medium, speckle light field signals output are interfered with reference light, and a speckle interference pattern is obtained;
[0008] S3, off-axis holographic calculation is performed on the speckle interference pattern obtained in step S2, and a speckle complex amplitude field transmitted via the scattering medium is constructed;
[0009] S4, an effective part of the speckle complex amplitude field obtained in step S3 is selected, the effective part is cropped to the same size as the multilayer phase information, and a training set and a test set are divided;
[0010] S5, the speckle complex amplitude field in step S3 is taken as a network input feature, a multilayer phase information reconstruction network is constructed, a complex domain calculation unit is introduced into the architecture of the multilayer phase information reconstruction network, so that the network can represent and process feature information in a complex value space, a complex domain loss function is defined to capture the error between predicted and actual axial multilayer phase information, and then the back propagation of the gradient is guided, the network parameters are iteratively optimized according to the training set and the test set, and finally a trained complete network is obtained to complete the reconstruction of the axial multilayer phase information.
[0011] Preferably, in step S1, the multilayer phase information includes three categories of two layers, three layers and four layers.
[0012] Preferably, in step S1, the axial relative distance of each layer of phase includes three categories of 0.001 m, 0.01 m and 0.1 m.
[0013] Preferably, in step S2, the method for acquiring the speckle complex amplitude field from the speckle interferogram adopts off-axis holography technology. The specific hologram reconstruction method is as follows: first, the acquired speckle interferogram is converted into its frequency domain using a two-dimensional Fourier transform, the autocorrelation term in the spectrum is filtered out, and one of the cross-correlation terms is moved to the center of the spectrum. Then, the spectrum plane is zero-filled, and the entire spectrum plane is inversely transformed with a two-dimensional Fourier transform, so as to obtain the distribution of the spatial domain complex amplitude field of the measured light field.
[0014] Preferably, in step S2, the superpixel encoding is to form a superpixel by combining multiple micromirrors on a digital micromirror device, and by designing the spatial distribution of their opening and closing, these micromirrors produce interference in the Fourier optical system, thereby synthesizing a complex amplitude field with specific amplitude and phase on the target plane.
[0015] Preferably, in step S3, the scattering medium includes frosted glass and multimode optical fiber.
[0016] Preferably, in step S5, the multi-layer phase information reconstruction network includes: complex domain convolution, complex domain activation function, complex domain transposed convolution, complex domain pooling, complex domain dense layer, complex domain batch normalization; the model architecture includes an encoder, a decoder, a jump connection, and an output layer, wherein the encoder part is used to gradually reduce the spatial dimension while increasing the number of feature channels, the decoder is used to upsample the feature map to the original image resolution, the jump connection is used to link the encoder features to the decoder of the corresponding resolution level, and the output layer is used to generate the output of the multi-layer reconstruction information.
[0017] Preferably, in step S5, the multi-layer phase information reconstruction network has one channel as input, and the input information is a speckle complex amplitude field containing speckle amplitude and phase information calculated from the speckle interferogram. A model is used to reconstruct the axially distributed multi-layer phase information, and the output is the complex amplitude field of multiple channels. The phase solved for the complex amplitude field output by each channel is the reconstructed phase information of each layer.
[0018] Compared with the prior art, the present invention has the following beneficial effects:
[0019] 1. The present invention adopts an end-to-end modeling approach, avoiding the complex intermediate steps of traditional methods, reducing the dependence on various preconditions, and greatly simplifying the complexity of the reconstruction task.
[0020] 2. The present invention adopts a deep neural network as a reconstruction method. Under large-scale training, it can not only adaptively learn the feature mapping between input and output, but also has good anti-interference ability, and has high scalability, versatility and robustness.
[0021] 3. This invention uses a complex-valued neural network to process speckle interferograms obtained directly from imaging, reconstructing multi-layer phase information including axial depth structure. This method uses the speckle complex amplitude field as the network input, providing not only richer prior information than traditional intensity maps but also incorporating complex domain computational units, significantly improving the model's ability to model the complex spatial patterns and depth-phase relationships in the speckle field. This allows for higher-precision reconstruction of multi-layer phase information across axial distance distributions in scattering media. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 This is a flowchart of the axial multi-layer phase information reconstruction method based on complex-valued neural network of the present invention.
[0023] Figure 2 This is a flowchart of a specific implementation case of the axial multi-layer phase information reconstruction method based on a complex-valued neural network of the present invention. DETAILED DESCRIPTION
[0024] The present invention will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present invention are not limited thereto.
[0025] See also Figure 1 The axial multi-layer phase information reconstruction method based on complex-valued neural network of the present invention comprises the following steps:
[0026] S1. Use MATLAB simulation software to simulate and calculate the complex amplitude field obtained after light passes through multiple layers of phase information placed at different relative distances in the axial direction;
[0027] S2. Based on the complex amplitude field obtained in step S1, holographic modulation is performed on it using superpixel coding, and then the signal is transmitted through a scattering medium. The output speckle light field signal interferes with the reference light to obtain a speckle interferogram.
[0028] S3, performing off-axis holographic calculation on the speckle interferogram obtained in step S2 to construct a complex amplitude field of the speckle after transmission through the scattering medium;
[0029] S4. Select the effective part of the speckle complex amplitude field obtained in step S3, crop it to the same size as the multi-layer phase information, and divide it into a training set and a test set;
[0030] S5, taking the speckle complex amplitude field described in step S3 as a network input feature, constructing a multi-layer phase information reconstruction network, introducing a complex domain calculation unit in the architecture of the multi-layer phase information reconstruction network, so that the network can represent and process feature information in the complex value space, defining a complex domain loss function to capture the error between the predicted and actual axial multi-layer phase information, and then guiding the back propagation of the gradient, iteratively optimizing the network parameters according to the training set and the test set, and finally obtaining a trained complete network to complete the reconstruction of the axial multi-layer phase information.
[0031] Referring to Figure 2 , the traditional method mainly uses the intensity speckle pattern as prior information to restore the target information, while in the embodiment, the speckle complex amplitude field is used as prior information, which can better reflect the depth relationship between the multi-layer phase information at different axial distances, so that the neural network can more easily learn the mapping relationship between the speckle field and the multi-layer phase information.
[0032] Referring to Figure 2 , the multi-layer phase information reconstruction network used in the embodiment includes: complex domain convolution, complex domain activation function, complex domain transposed convolution, complex domain pooling, complex domain dense layer, and complex domain batch normalization; the model architecture includes an encoder, a decoder, a skip connection, and an output layer.
[0033] Referring to Figure 2 , the loss function of the multi-layer phase information reconstruction network is an MSE function, which calculates the average of the absolute values between the predicted value and the true value, and is used in the network to compare the similarity between the multi-channel output phase and the multi-layer label phase. The calculation method of the loss function is to calculate the average of the loss function values between each corresponding channel of the output phase and each layer of the label phase.
[0034] Through the above settings, the axial multi-layer phase information reconstruction method based on the complex value neural network uses the speckle complex amplitude field as the network input to obtain better multi-layer phase reconstruction effect than using the speckle intensity as the input, and when facing the 2-4 layer phase information reconstruction with axial phase distance of 0.001m, 0.01m and 0.1m, the multi-layer phase information reconstructed by the method also has high accuracy.
[0035] The above is the preferred embodiment of the present application, but the embodiments of the present application are not limited by the above, any change, modification, replacement, combination, simplification made without departing from the spirit and principles of the present application shall be equivalent replacement mode, all included in the protection scope of the present application.
Claims
1. A method for reconstructing axial multi-layer phase information based on a complex-valued neural network, characterized in that: The method comprises the following steps: S1. Use MATLAB simulation software to simulate and calculate the complex amplitude field obtained after light passes through multiple layers of phase information placed at different relative distances in the axial direction; S2. Based on the complex amplitude field obtained in step S1, holographic modulation is performed on it using superpixel coding, and then the signal is transmitted through a scattering medium. The output speckle light field signal interferes with the reference light to obtain a speckle interferogram. S3, performing off-axis holographic calculation on the speckle interferogram obtained in step S2 to construct a complex amplitude field of the speckle after transmission through the scattering medium; S4. Select the effective part of the speckle complex amplitude field obtained in step S3, crop it to the same size as the multi-layer phase information, and divide it into a training set and a test set; S5. Using the speckle complex amplitude field described in step S3 as a network input feature, a multi-layer phase information reconstruction network is constructed. A complex domain computing unit is introduced into the architecture of the multi-layer phase information reconstruction network, so that the network can represent and process feature information in a complex-valued space. By defining a complex domain loss function, the error between the predicted and actual axial multi-layer phase information is captured, thereby guiding the back propagation of the gradient. The network parameters are iteratively optimized based on the training set and the test set, and finally a fully trained network is obtained to complete the reconstruction of the axial multi-layer phase information.
2. The axial multi-layer phase information reconstruction method based on complex-valued neural network according to claim 1 is characterized in that: In step S1, the multi-layer phase information includes three categories: two layers, three layers, and four layers.
3. The axial multi-layer phase information reconstruction method based on complex-valued neural network according to claim 1 is characterized in that: In step S1 , the axial relative distance of each phase layer includes three categories: 0.001m, 0.01m, and 0.1m.
4. The axial multi-layer phase information reconstruction method based on complex-valued neural network according to claim 1 is characterized in that: In step S2, the superpixel encoding used is to combine multiple micromirrors on a digital micromirror device into a superpixel. By designing the spatial distribution of their opening and closing, these micromirrors produce interference in the Fourier optical system, thereby synthesizing a complex amplitude field with specific amplitude and phase on the target plane.
5. The method for reconstructing axial multi-layer phase information based on a complex-valued neural network according to claim 1, characterized in that: In step S2, the method for acquiring the speckle complex amplitude field from the speckle interferogram adopts off-axis holography technology. The specific hologram reconstruction method is as follows: first, the acquired speckle interferogram is converted into its frequency domain using a two-dimensional Fourier transform, the autocorrelation terms in the spectrum are filtered out, and one of the cross-correlation terms is moved to the center of the spectrum. Then, the spectrum plane is zero-filled, and the entire spectrum plane is subjected to a two-dimensional inverse Fourier transform. In this way, the distribution of the spatial domain complex amplitude field of the measured light field can be obtained.
6. The method for reconstructing axial multi-layer phase information based on a complex-valued neural network according to claim 1, characterized in that: In step S2, the scattering medium includes frosted glass and multimode optical fiber.
7. The axial multi-layer phase information reconstruction method based on complex-valued neural network according to claim 1 is characterized in that: In step S5, the multi-layer phase information reconstruction network includes: complex domain convolution, complex domain activation function, complex domain transposed convolution, complex domain pooling, complex domain dense layer, and complex domain batch normalization; the model architecture includes an encoder, a decoder, a jump connection, and an output layer, wherein the encoder part is used to gradually reduce the spatial dimension while increasing the number of feature channels, the decoder is used to upsample the feature map to the original image resolution, the jump connection is used to link the encoder features to the decoder of the corresponding resolution level, and the output layer is used to generate the output of the multi-layer reconstruction information.
8. The method for reconstructing axial multi-layer phase information based on a complex-valued neural network according to claim 5, characterized in that: For the multi-layer phase information reconstruction network, the input is a channel, and the input information is the speckle complex amplitude field containing speckle amplitude and phase information calculated from the speckle interferogram. A model is used to reconstruct the axially distributed multi-layer phase information. Its output is the complex amplitude field of multiple channels. The phase solved for the complex amplitude field output by each channel is the reconstructed phase information of each layer.