Blind single phase recovery method based on self-supervised physical adaptive network
Through the multi-stage coupled reconstruction framework and fine loss function of self-supervised physical adaptive network, the robustness problem of blind phase recovery under large defocus distance is solved, and high-quality image reconstruction is achieved.
Patent Information
- Application Number
- CN202510556609.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-29
- Publication Date
- 2025-08-08
AI Technical Summary
The existing blind phase recovery method is poorly robust under large defocus distance errors, making it difficult to achieve high-quality image reconstruction.
A self-supervised physical adaptive network is adopted, through multi-stage coupled reconstruction framework and fine physical consistent loss function, jointly optimize physical parameters and image reconstruction, and a micro-punishment mechanism is introduced to guide the training process to realize self-supervised training.
High-quality image reconstruction is achieved under large diffraction distance error, improving stability and robustness, and the image reconstruction quality and robustness are better than existing methods.
Smart Images
Figure CN120451017A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a phase recovery technology in the field of optical imaging technology, and in particular to a blind single-shot phase recovery method based on a self-supervised physical adaptive network. Background Art
[0002] Phase retrieval represents a fundamental challenge in optics and is an integral part of modern computational imaging systems, including but not limited to coherent diffraction imaging, Fourier plane microscopy, computational holography, and optical metrology. The goal of phase retrieval is to reconstruct the original sample information captured by the detector through efficient information encoding and robust algorithms.
[0003] Phase retrieval is inherently a nonconvex and nonlinear inverse problem, presenting considerable computational challenges. To address these challenges, a range of advanced imaging modalities have been proposed, including coded diffraction patterns, multi-range measurements, and plane scanning. However, these approaches often increase experimental complexity or impose additional physical constraints, often requiring the acquisition of multiple-intensity images. This necessity limits their applicability in dynamic or in vivo imaging scenarios. To address these challenges, single-shot phase retrieval techniques have garnered significant attention due to their ability to reconstruct a high-fidelity sample representation from a single-intensity measurement.
[0004] While existing single-frame phase retrieval algorithms demonstrate impressive reconstruction performance under ideal imaging conditions, they rely heavily on accurate physical forward models and are highly sensitive to parameters such as defocus distance. In real-world imaging scenarios, factors such as mechanical drift of the imaging platform or displacement of the 3D sample focal plane often lead to unknown or inaccurately measured physical propagation distances. To address this issue, many systems incorporate an additional autofocus step to calibrate the defocus distance, which limits both flexibility and accuracy.
[0005] In recent years, blind network frameworks have accounted for the uncertainty in defocus distance and solved phase recovery under unknown defocus conditions. However, joint optimization of physical parameters and image reconstruction still faces significant challenges, especially when large defocus distance errors are involved. This emphasizes the need for robust and adaptive methods that can simultaneously improve physical parameter estimation and image reconstruction under non-ideal conditions. Summary of the Invention
[0006] The present invention aims to solve the problem of poor robustness of existing blind phase recovery methods under large defocus distance errors, and provides a blind single-shot phase recovery method based on a self-supervised physical adaptive network, which is widely used in the field of optical imaging, such as computational microscopy and holographic imaging.
[0007] Self-supervised blind phase recovery is accomplished using a self-supervised physically adaptive neural network. This method employs a novel network architecture that integrates adjustable physical parameters into a multi-stage coupled reconstruction process. Through self-supervised network training facilitated by a sophisticated physically consistent loss function, the physical parameters of the feedforward model are jointly optimized with the trainable parameters of the reconstruction network, achieving high-quality image reconstruction despite large diffraction distance errors.
[0008] In order to achieve the above object of the invention, the technical solution adopted by the present invention is as follows:
[0009] The blind single-shot phase recovery method based on a self-supervised physical adaptive network includes the following steps:
[0010] S1: Establish a self-supervised blind phase recovery method based on the coherent diffraction system in optical propagation theory, and jointly optimize the physical parameters and target image;
[0011] S2: According to the physical propagation model based on the angular spectrum method, the optimized target image is converted into the optimization of network parameters, and the corresponding single-shot reconstructed image and the corresponding diffraction distance are obtained by simulation;
[0012] S3: Establish a physical adaptive network, take the collected real diffraction pattern as input, convert the phase information into physical parameters, and realize the reconstruction of blind images and accurate estimation of physical parameters.
[0013] S4: Establish a refined physical consistent loss function and self-supervised training model, jointly constrain the optimization of physical parameters and network parameters through detail loss and structure loss, and introduce a micro-penalty mechanism to guide training.
[0014] The real diffraction pattern captured by the coherent diffraction system is used as input and the blind phase recovery algorithm is used to reconstruct the image.
[0015] Furthermore, S1 established a blind phase recovery method based on self-supervision, which jointly optimized the physical parameters and the target image as follows:
[0016] According to the physical parameter self-correction method, the regularized optimization problem is solved. The model can be expressed as:
[0017]
[0018] in, is the reconstructed image, y is the real diffraction pattern, H d (·) is a physical propagation model based on the angular spectrum method, d is the diffraction distance, is a regularization term, and β is a hyperparameter used to control the weight of the regularization term.
[0019] Considering that the above formula requires autofocusing on the diffraction distance, the regularized optimization problem is converted into a problem of jointly optimizing physical parameters and the target image to achieve blind single-shot phase recovery:
[0020]
[0021] In S2, in order to improve the representation ability of the reconstructed image and reduce the optimization iteration time, an untrained neural network structure is used as a regularizer to convert the optimization target image into the optimization of the network parameters. The model is constructed as follows:
[0022]
[0023] Among them, f θ (y) represents a neural network parameterized by θ that maps the diffraction pattern y back to the sample image x.
[0024] To prevent the network from falling into a local optimal solution during training, the adjustable physical parameters are integrated with the network parameters:
[0025]
[0026] Among them, Θ is the trainable parameter of the entire network.
[0027] In S3, a physical adaptive network is proposed in network construction, and a multi-stage reconstruction framework is established. Each stage mainly includes two parts: physical model update and image regularization update.
[0028] For the physical model update part of the kth stage, based on the propagation distance parameter provided by the current network parameter Θ1, the forward propagation model is used The reconstructed image is iteratively refined using analytical gradient descent:
[0029]
[0030] Among them, x k Represents the current reconstructed image, r k Indicates the step size.
[0031] For the image regularization update part of the kth stage, a convolutional neural network based on the UNet architecture is used. In order to further enhance the network's ability to represent images, the channel attention (CA) mechanism is embedded into the convolutional neural network:
[0032]
[0033] where x k+1 represents the output image of the image regularization update part of the current stage, and Θ2 is the network parameter of the embedded channel attention UNet.
[0034] In S4, in the loss function part, a sophisticated physically consistent loss function and self-supervised training model are established.
[0035] The fine physical consistency loss function consists of two parts: detail loss and structure loss.
[0036] The loss of detail is quantified by the peak signal-to-noise ratio (PSNR) between the predicted and observed diffraction patterns, while the loss of structure is evaluated by the structural similarity index measure (SSIM).
[0037] Since the present invention only uses the collected diffraction patterns without real images as reference, the entire training process is self-supervised.
[0038] The present invention introduces a conditional micro-penalty mechanism to dynamically adjust and guide the training process, ensuring enhanced convergence and improved robustness.
[0039] Compared with the prior art, the advantages of the present invention are:
[0040] The present invention proposes a multi-level network architecture, which integrates adjustable physical parameters into the coupled reconstruction framework, promotes the joint optimization of the physical parameters controlled by the forward propagation model and the trainable parameters of the reconstruction network, and solves the problem of blind single-frame phase recovery.
[0041] We introduce a sophisticated physically consistent loss function to enable self-supervised training, ensuring excellent stability and robustness, especially in cases involving large diffraction range errors.
[0042] The present invention can reconstruct the original image from only a single observed diffraction pattern, and the reconstruction quality (measured by image peak signal-to-noise ratio) and algorithm robustness are much higher than those of existing blind single-frame phase recovery methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 A device for collecting real diffraction images used in the present invention;
[0044] Figure 2 This is a real diffraction image collected in the embodiment;
[0045] Figure 3 Reconstructed image for the embodiment. DETAILED DESCRIPTION
[0046] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples.
[0047] The blind single-shot phase recovery method based on self-supervised physical adaptive network mainly consists of two parts, including a blind phase recovery method based on coherent diffraction system and a self-supervised physical adaptive network.
[0048] First, the phase retrieval technology based on traditional coherent diffraction imaging was improved to jointly optimize physical parameters and target images, which solved the blind phase retrieval problem and realized the reconstruction of the diffraction image.
[0049] Traditional phase retrieval technology based on coherent diffraction systems uses a method of physical parameter self-correction to reconstruct the diffraction image:
[0050]
[0051] in, is the reconstructed image, y is the real diffraction pattern, H d (·) is a physical propagation model based on the angular spectrum method, d is the diffraction distance, is a regularization term, and β is a hyperparameter used to control the weight of the regularization term.
[0052] The regularized optimization problem based on physical parameter self-correction in the above formula is improved into a joint optimization problem:
[0053]
[0054] In order to improve the representation ability of the reconstructed image and reduce the number and time of optimization iterations, an untrained neural network structure is used as a regularizer to convert the optimization target image into the optimization of network parameters. The model is constructed as follows:
[0055]
[0056] Among them, f θ (y) represents a neural network parameterized by θ that maps the diffraction pattern y back to the sample image x.
[0057] To prevent the network from falling into a local optimal solution during training, the adjustable physical parameters are integrated with the network parameters:
[0058]
[0059] Among them, Θ is the trainable parameter of the entire network.
[0060] After the modeling is completed, Figure 1 The collected coherent diffraction image y is used as the network input, as shown in Figure 2 As shown in the figure, its size is 1080*1920, and x1 is initialized to all 1s and the diffraction distance is Θ1 to 6 mm. The reconstructed image and the accurate diffraction distance are obtained through the physical adaptive network.
[0061] First, input the part updated by the physical model, and at the kth stage, get the current output image:
[0062]
[0063] Among them, v k Update partial output for the current stage of the physical model, r k Indicates that the step size is set to 1.
[0064] Then, through the image regularization update part, the current output image is obtained and the network parameters are updated:
[0065]
[0066] Among them, x k+1 Update part of the output for the image regularization of the current stage, that is, the output of the physical adaptation network of the kth stage.
[0067] Based on the current diffraction distance, the network output is degraded:
[0068]
[0069] Among them, y p is the degraded image at the current stage.
[0070] Calculate the mean square error and set the loss function threshold T = 10000:
[0071]
[0072] Detail loss and structure loss are calculated. Detail loss is quantified by the peak signal-to-noise ratio (PSNR) between the predicted and observed diffraction patterns, while structure loss is evaluated using the structural similarity index measure (SSIM).
[0073] Calculate the loss function and introduce a conditional micro-penalty mechanism to dynamically adjust and guide the training process:
[0074] loss=MSE / PSNR (9)
[0075]
[0076] Update the diffraction image x k The diffraction distance is Θ1, and the loss function threshold T = loss is updated.
[0077] In this way, the blind single-shot phase recovery method based on the self-supervised physical adaptive network can be obtained by iterating according to equations (5)-(10). Figure 3 The reconstructed image is shown.
[0078] The key advantage of this invention lies in its proposed multi-level network architecture, which integrates adjustable physical parameters into a coupled reconstruction framework. This facilitates the joint optimization of the physical parameters controlled by the forward propagation model and the trainable parameters of the reconstruction network, solving the problem of blind single-frame phase recovery. Furthermore, a sophisticated physically consistent loss function is introduced to enable self-supervised training, ensuring excellent stability and robustness, especially in situations involving large diffraction distance errors.
Claims
1. A blind single-shot phase recovery method based on a self-supervised physical adaptive network, characterized by: The following steps are involved: S1: Establish a self-supervised blind phase recovery method based on the coherent diffraction system in optical propagation theory, and jointly optimize the physical parameters and target image; S2: According to the physical propagation model based on the angular spectrum method, the optimized target image is converted into the optimization of network parameters, and the corresponding single-shot reconstructed image and the corresponding diffraction distance are obtained by simulation; S3: Establish a physical adaptive network, take the collected real diffraction pattern as input, convert the phase information into physical parameters, and realize the reconstruction of blind images and accurate estimation of physical parameters; S4: Establish a refined physical consistent loss function and self-supervised training model, jointly constrain the optimization of physical parameters and network parameters through detail loss and structure loss, and introduce a micro-penalty mechanism to guide training.
2. The blind single-shot phase recovery method based on a self-supervised physical adaptive network according to claim 1, characterized in that The specific method of S1 is: A self-supervised blind phase recovery method is established to reconstruct the true diffraction pattern by solving a regularized optimization problem: Among them, x is the initialized all-1 image, is the reconstructed image, y is the real diffraction pattern, H d (·) is a physical propagation model based on the angular spectrum method, d is the diffraction distance, is a regularization term, and β is a hyperparameter used to control the weight of the regularization term; Considering that Equation (1) requires autofocus diffraction distance, the regularized optimization problem is converted into a problem of jointly optimizing physical parameters and target image to achieve blind single-shot phase recovery:
3. The blind single-shot phase recovery method based on self-supervised physical adaptive network according to claim 2, characterized in that: In S2, the optimization target image is converted into the optimization of network parameters. The specific method is: Use an untrained neural network structure as a regularizer to convert the optimization image into trainable network parameters: Among them, f θ (y) represents a neural network parameterized by θ that maps the diffraction pattern y back to the sample image x. In order to avoid local optimal solutions during network training, the adjustable physical parameters are integrated with the network parameters: Among them, Θ is the trainable parameter of the entire network.
4. The blind single-shot phase recovery method based on self-supervised physical adaptive network according to claim 3 is characterized in that Establish a physical adaptive network in S3. The specific method is as follows: A multi-stage reconstruction framework is proposed, where each stage includes two parts: physical model update and image regularization update. For the physical model update part of the kth stage, based on the propagation distance parameter provided by the current network parameter Θ1, the forward propagation model is used The reconstructed image is iteratively refined using the analytical gradient descent method as follows: Among them, x k Represents the current reconstructed image, r k Indicates the step length; For the image regularization update part of the kth stage, a convolutional neural network based on the UNet architecture is used. In order to enhance the representation ability of the network, the channel attention CA mechanism is embedded into the convolutional neural network: where x k+1 represents the output image of the image regularization update part of the current stage, and Θ2 is the network parameter of the embedded channel attention UNet.
5. The blind single-shot phase recovery method based on self-supervised physical adaptive network according to claim 1, characterized in that In S4, a refined physically consistent loss function and self-supervised training model are established. The specific method is as follows: It is proposed to jointly construct the loss function by detail loss and structure loss; The detail loss is quantified by the peak signal-to-noise ratio (PSNR) between the predicted and observed diffraction patterns, and the structure loss is evaluated by the structural similarity index (SSIM). Only the collected diffraction patterns are used, without real images as reference, and the entire training process is self-supervised; A conditional micro-penalty mechanism is introduced to dynamically adjust and guide the training process to ensure enhanced convergence and improved robustness.