Strong-turbulence phase-shifting interferometry phase retrieval system and method based on spatiotemporal deep network

Abstract

The present application relates to the technical field of optical detection, and particularly relates to a strong turbulent phase shift interferometric phase retrieval system and method based on a space-time deep network, statistical characteristics of an acquired phase shift interferometric image sequence; time direction information extraction and attention enhancement are performed on the phase shift interferometric image sequence to obtain time attention characteristics and time statistical characteristics; space fusion characteristics composed of the phase shift interferometric image sequence and the statistical characteristics are fused with the time attention characteristics and the time statistical characteristics in a space direction to obtain space-time joint characteristics; encoding and decoding operations are performed on the space-time joint characteristics to obtain multi-scale phase characteristics; and the multi-scale phase characteristics are respectively recovered to obtain final output phases. The present application does not use physical initial values to generate initial phases, and directly learns phase mapping relationships from multiple frames of interferometric images by a deep network, so that the accuracy, stability and anti-interference ability of phase shift interferometric phase retrieval can be significantly improved under strong turbulent conditions.

Description

Technical Field

[0001] This invention belongs to the field of optical detection technology, and particularly relates to a strong turbulent phase shift interference phase recovery system and method based on spatiotemporal depth networks. Background Technology

[0002] Phase-shift interferometry is a key technology in optical surface inspection, wavefront detection, precision manufacturing quality control, and high-resolution imaging. Its basic idea is to acquire interferometric images at multiple known phase-shift positions, establish a mapping relationship between phase and light intensity through intensity changes, and then obtain the phase distribution to be measured through phase demodulation. This type of method has advantages such as non-contact operation, high sensitivity, and high resolution, and has been widely used in precision optics, semiconductor inspection, micrometrology, and industrial online inspection.

[0003] Existing phase-shifting interferometric phase retrieval methods mostly employ analytical or iterative demodulation strategies, such as window function-weighted phase retrieval methods, weighted polynomial-based phase retrieval methods, and least-squares fitting-based phase estimation methods. These methods achieve good results under ideal imaging conditions, stable phase shifts, and high signal-to-noise ratios. However, in practical applications, they are often affected by the following factors: First, traditional analytical methods are sensitive to outliers due to light source fluctuations, sensor noise, non-uniform modulation, local occlusion, or low-reflection areas. Second, temporal correlations exist between multiple frames, making it difficult for traditional pixel-by-pixel analytical methods to effectively utilize cross-frame redundancy information. Third, while data-driven deep learning methods have strong fitting capabilities, they often lack physical constraints, leading to unstable results, insufficient generalization, or weak interpretability.

[0004] Therefore, how to construct a phase retrieval method that does not rely on analytical physical initial values ​​and can make full use of multi-frame temporal information and spatial context features has become a key issue in improving the accuracy and robustness of strong turbulent phase shift interferometry. Summary of the Invention

[0005] In view of this, the present invention aims to provide a strong turbulent phase shift interferometric phase recovery system and method based on spatiotemporal deep networks. It does not use physical initial values ​​to generate the initial phase, but learns the phase mapping relationship directly from multiple frames of interferometric images by the deep network. It can significantly improve the accuracy, stability and anti-interference ability of phase shift interferometric phase recovery under strong turbulent conditions.

[0006] To achieve the above objectives, the technical solution created by this invention is implemented as follows: A phase retrieval system for strong turbulent phase-shift interferometric imaging based on a spatiotemporal deep network includes: a frame statistics unit, which performs statistical calculations on the acquired phase-shift interferometric image sequence to obtain statistical features; a temporal modeling unit, which extracts temporal information and enhances attention on the input image stack composed of the phase-shift interferometric image sequence to obtain temporal attention features and temporal statistical features; a spatial modeling unit, which fuses the spatial fusion features composed of the phase-shift interferometric image sequence and statistical features with the temporal attention features and temporal statistical features in a spatial direction to obtain spatiotemporal joint features; a phase extraction unit, which performs hierarchical encoding and decoding operations on the spatiotemporal joint features to obtain multi-scale phase features; and a phase retrieval unit, which recovers the phase features at each of the multi-scales, with the result of recovering the phase shift feature at the largest scale being the final output phase.

[0007] Furthermore, in the temporal modeling unit: the input image stack is expanded in dimension, and then the expanded features are subjected to 3D convolution to obtain preliminary features; the preliminary features are subjected to 3D convolution and softmax operations to obtain temporal attention weights, and then the temporal attention weights are attached to the preliminary features to obtain temporal attention fusion features; the temporal attention fusion features are subjected to time mean and time maximum response calculation to obtain temporal statistical features.

[0008] Furthermore, in the spatial modeling unit: spatial fusion features are convolved in two dimensions to obtain spatial statistical features; spatial statistical features, temporal attention features, and temporal statistical features are concatenated and convolved to obtain spatiotemporal joint features.

[0009] Furthermore, in the phase extraction unit, a U-Net network is used to perform step-by-step encoding and decoding operations on the spatiotemporal joint features to obtain multi-scale phase features.

[0010] Furthermore, the phase recovery unit includes an output head corresponding to the phase feature scale; each output head maps the phase feature of the corresponding scale to a two-channel phase representation feature, one channel being a cosine component and the other channel being a sine component; after normalizing the phase representation feature, the phase is obtained through the arctangent function.

[0011] A phase retrieval method for strong turbulent phase shift interferometry based on spatiotemporal deep networks includes: S1: Preprocess the acquired multiple sets of phase-shifting interferometric image sequences and the corresponding real phase maps to obtain the training set; S2: Construct a strong turbulent phase shift interference phase recovery system based on spatiotemporal depth network as provided in this invention; S3: Use the training set built in step S1 to train the system built in step S2 to obtain the phase retrieval model; S4: Input the phase-shifting interferometric image sequence to be tested into the phase recovery model obtained in step S3, and output the final phase map.

[0012] Furthermore, in step S3, the system is trained using the following joint loss function: L total =λ1L ph +λ2L sc +λ3L grad +λ4L ms +λ5L smooth +λ6L edge ; Among them, L total L represents the joint loss function. ph L represents phase loss. sc L represents the sine and cosine consistency loss. grad L represents the gradient loss. ms L represents multi-scale loss. smooth L represents the smoothing loss, used to suppress ineffective high-frequency oscillations. edge λ1~λ6 represent the edge consistency loss, and λ1~λ6 represent the loss weights; The phase loss is: ; Among them, L Huber Represents the Huber sub-loss of the fundamental phase, η i This represents the Huber sub-loss weight corresponding to the phase map at the i-th scale before unifying the scale. This represents the phase diagram at the i-th scale before scaling. Represents the true phase diagram; The sine and cosine consistency loss is: ; Where Ω represents the non-zero region in the true phase diagram, and |Ω| represents the total number of elements in the non-zero region. This represents the phase diagram that the system finally recovers. This represents the corresponding true phase diagram; The gradient loss is: ; in, and These represent the calculation of the first-order gradient in the horizontal and vertical directions, respectively. The smoothing loss is: ; The edge consistency loss is: ; in, This represents the calculation of the second-order Laplace derivative; The multi-scale loss is: ; in, This represents the phase map after downsampling at the j-th mesoscale from the phase map finally recovered by the system. This represents the phase map after downsampling the true phase map at the j-th mesoscale, γ. j This represents the corresponding multi-scale loss weight.

[0013] Furthermore, the basic phase Huber sub-loss is: ; Where Huber represents the circumferential phase Huber loss, which is: ; Where ε represents the phase diagram Compared with the true phase diagram The phase difference between them, δ represents the supervision threshold.

[0014] Compared with the prior art, the present invention can achieve the following beneficial effects: This invention presents a strong turbulent phase shift interferometric phase recovery system and method based on spatiotemporal deep networks. It does not rely on analytical physical phase initial values, thus avoiding error propagation caused by erroneous priors under strong turbulent conditions. Through three-dimensional temporal convolution and temporal attention aggregation, it can fully utilize the dynamic correlation between multiple phase-shifted images. Through frame statistical feature branching and U-shaped multi-scale backbone collaborative optimization, it can simultaneously improve the global phase structure recovery capability and the local edge detail preservation capability. Through multi-scale deep supervision and composite loss constraints, it still maintains high stability and recovery accuracy under conditions of strong turbulence, low-level conditions, and outliers. Attached Figure Description

[0015] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic diagram of the overall turbulent phase-shift interferometric phase recovery system based on a spatiotemporal depth network as described in an embodiment of the present invention; Figure 2 A schematic flowchart of the phase recovery method for strong turbulent phase shift interference based on spatiotemporal depth networks, as described in an embodiment of the present invention; Figure 3 The RMSE comparison result diagram described in the embodiment of the present invention; Figure 4 The MAE comparison result diagram described in the embodiment of the present invention; Figure 5 The graph showing the change of the loss function during the model training process described in the embodiments of the present invention; Figure 6 A graph showing the change in the root mean square error of the phase during the model training process described in the embodiments of the present invention; Figure 7 The graph showing the change in learning rate during the model training process described in the embodiments of the present invention. Detailed Implementation

[0016] 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 specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0017] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] like Figure 1 As shown in the embodiment of the present invention, the strong turbulent phase-shift interferometric phase retrieval system based on a spatiotemporal deep network includes a frame statistics unit, a temporal modeling unit, a spatial modeling unit, a phase extraction unit, and a phase retrieval unit. The frame statistics unit performs statistical calculations on the acquired phase-shift interferometric image sequence to obtain statistical features; the temporal modeling unit extracts temporal information and enhances attention on the input image stack composed of the phase-shift interferometric image sequence to obtain temporal attention features and temporal statistical features; the spatial modeling unit fuses the spatial fusion features composed of the phase-shift interferometric image sequence and statistical features with the temporal attention features and temporal statistical features in a spatial direction to obtain spatiotemporal joint features; the phase extraction unit performs hierarchical encoding and decoding operations on the spatiotemporal joint features to obtain multi-scale phase features; the phase retrieval unit recovers the multi-scale phase features respectively, with the result of recovering the phase shift feature with the largest scale being the final output phase.

[0019] In this embodiment of the invention, the phase-shifting interferometric image sequence includes 16 phase-shifting interferometric images in chronological order. A frame statistics unit calculates the frame mean, frame standard deviation, and frame maximum value of the 16 phase-shifting interferometric images, performing statistical calculations. The resulting frame mean, frame standard deviation, and frame maximum value constitute statistical features. These statistical features can reflect the degree of observational degradation of the phase-shifting interferometric image under strong turbulence without an explicit physical model. A spatial modeling unit performs channel stitching on the image sequence composed of the 16 phase-shifting interferometric images and the statistical features including the frame mean, frame standard deviation, and frame maximum value, resulting in a 19-dimensional spatial fusion feature used to characterize local brightness distribution, fluctuation amplitude, and high-response regions.

[0020] In some embodiments, the feature processing in the temporal modeling unit includes: expanding the dimensions of the input image stack to construct the data structure format required for subsequent 3D convolution, avoiding the temporal dimension from being mistakenly treated as channel mixing, and providing a structural basis for subsequent processing; performing 3D convolution on the expanded features to obtain preliminary features; performing 3D convolution and softmax operations on the preliminary features to obtain temporal attention weights, and then attaching the temporal attention weights to the preliminary features to obtain temporal attention fusion features; calculating the temporal mean and temporal maximum response of the temporal attention fusion features to obtain temporal statistical features. Specifically, the temporal mean calculation involves calculating the mean of the temporal attention fusion features along the time dimension. The purpose of this calculation is to extract stable structures, remove turbulent random fluctuations, and retain long-term consistent information. The temporal maximum response calculation involves taking the maximum value of the temporal attention fusion features along the time dimension. The purpose of this calculation is to emphasize instantaneous peak information and counteract the clear structure of a particular frame in turbulence.

[0021] The obtained temporal statistical features include two types of information: temporal mean and temporal maximum response. Combined with temporal attention fusion features, the three together characterize the dynamic evolution information of the phase-shifting interferometric image sequence under strong turbulence conditions.

[0022] In this embodiment of the invention, the input image stack (i.e., 16 phase-shifting interferometric images organized in chronological order) has dimensions [B,T,H,W], where B represents the input batch, T represents the temporal length, and H and W represent the height and width of the phase-shifting interferometric image, respectively. After expanding the dimensions, a tensor of [B,1,T,H,W] is formed. Spatiotemporal features are extracted sequentially through three layers of three-dimensional convolution, group normalization, and SiLU activation to obtain preliminary features. The preliminary features are then processed through a 1×1×1 convolution to generate temporal attention weights, and softmax normalization is performed in the temporal dimension to obtain normalized temporal attention weights.

[0023] In some embodiments, the feature processing in the spatial modeling unit includes: performing two-dimensional convolution on the spatial fusion features to obtain spatial statistical features; and performing channel concatenation and convolution on the spatial statistical features, temporal attention features, and temporal statistical features to obtain spatiotemporal joint features.

[0024] In this embodiment of the invention, the 19-dimensional spatial fusion features are output as spatial statistical features through two-dimensional convolution and residual blocks; after channel splicing and convolution processing of the spatial statistical features, temporal attention features and temporal statistical features, a spatiotemporal joint feature with 48 channels is obtained.

[0025] In some embodiments, the feature processing of the phase extraction unit includes: using a U-Net network to perform stepwise encoding and decoding operations on the spatiotemporal joint features to obtain multi-scale phase features.

[0026] In this embodiment of the invention, the encoding branch of the U-Net network includes a first-level residual feature block and a third-level downsampling module, with the number of channels progressively increasing from 48 to 96, 192, and 384. Each residual convolutional block employs group normalization, SiLU activation, optional Dropout, and SE channel attention units to improve training stability and feature representation capabilities. At the junction of the encoding and decoding branches of the U-Net network, a bottleneck layer is used for feature transition. This bottleneck layer consists of two levels of residual convolutional blocks and is used to further extract high-level semantic information. In the decoding branch of the U-Net network, a three-level upsampling structure is used to gradually restore spatial resolution, and high-resolution details from the encoding end are introduced through skip connections with a gating mechanism. The gating mechanism generates skip feature weights through 1×1 convolutions, which are then multiplied point-by-point with the skip connection features to suppress invalid textures and noise propagation. In this embodiment of the invention, the skip connection with gating mechanism is a prior art (Oktay O, Schlemper J, Folgoc LL, et al. Attention U-Net: Learning Where to Look for the Pancreas[J].2018.DOI:10.48550 / arXiv.1804.03999).

[0027] In some embodiments, the phase recovery unit includes an output head corresponding to the phase feature scale; each output head maps the phase feature of the corresponding scale to a two-channel phase representation feature, one channel being a cosine component and the other channel being a sine component; after normalizing the phase representation feature, the phase is obtained through the arctangent function.

[0028] In this embodiment of the invention, phase maps of three scales are output: fine-scale phase map, medium-scale phase map, and coarse-scale phase map. The phase feature with the largest scale is processed by the corresponding output head to obtain the fine-scale phase map. Similarly, the phase features with the medium and smallest scales are processed by the corresponding output heads to obtain the medium-scale phase map and the coarse-scale phase map, respectively. The fine-scale phase map is the phase map that is finally recovered.

[0029] This invention also provides a phase retrieval method for strong turbulent phase shift interferometry based on spatiotemporal depth networks, combined with... Figure 1 and Figure 2 ,include: S1: Preprocess the acquired multiple sets of phase-shifting interferometric image sequences and the corresponding real phase maps to obtain the training set.

[0030] In this embodiment of the invention, each set of phase-shifting interferometric image sequences includes 16 frames of phase-shifting interferometric images. The acquisition process of each set of phase-shifting interferometric image sequences is based on the method provided in the existing paper (Tang JU, Ji WU, Zhang J, et al. Highly robust spatiotemporal wavefront prediction with a mixed graph neural network in adaptive optics[J]. Photonics Research (English), 2023(11).), including: A turbulence physical model is constructed, which is based on the Kolmogorov / von Kármán turbulence theory. On this basis, a multi-layer phase screen and angular spectrum propagation are used to simulate the air refractive index fluctuations in the long optical path. At the same time, the Taylor frozen flow hypothesis is combined to introduce the AR(1) time series model to reproduce the dynamic airflow disturbances related to the frame and obtain the turbulence physical model. A standard phase-shifting interference optical path is constructed in the turbulence physics model to generate an ideal original interference light intensity distribution with no interference and regular fringes. Various systematic errors, such as phase shift step size deviation, Gaussian / Poisson imaging noise, motion vibration, image blurring and random bad frames, are introduced into the ideal original interference light intensity distribution to fit the actual measurement degradation effect and obtain the generation model of each frame of phase shift interference image; Multiple simulations were performed on the generative model to obtain multiple sets of phase-shifted interferometric image sequences, each consisting of 16 frames of phase-shifted interferometric images.

[0031] Furthermore, the data preprocessing process in this embodiment includes: uniformly adjusting the size of all frame phase-shifted interferometric images to 256×256 pixels. Additionally, the data preprocessing process in this embodiment includes: for the phase-shifted interferometric images, performing random horizontal flipping, random vertical flipping, and random transposition to improve the model's generalization ability to spatial distribution variations; for the true phase map, performing zero-mean normalization on the phase values ​​of non-zero regions in the true phase map to reduce the impact of global bias on training. Furthermore, to mitigate the adverse effects of global brightness drift on training, this embodiment performs frame-by-frame standardization processing on each frame of the phase-shifted interferometric image, specifically by subtracting the frame mean from the single-frame pixel value and dividing by the frame standard deviation.

[0032] S2: Construct a strong turbulent phase shift interference phase recovery system based on spatiotemporal depth network as provided in this invention.

[0033] S3: Use the training set built in step S1 to train the system built in step S2 to obtain the phase retrieval model.

[0034] In some embodiments, the system is trained using the following joint loss function: L total =λ1L ph +λ2L sc +λ3L grad +λ4L ms +λ5L smooth +λ6L edge ; Among them, L total L represents the joint loss function. ph L represents phase loss. sc L represents the sine and cosine consistency loss. grad L represents the gradient loss. ms L represents multi-scale loss. smooth L represents the smoothing loss, used to suppress ineffective high-frequency oscillations. edge λ represents the edge consistency loss, and λ1~λ6 represent the loss weights.

[0035] The phase loss is: ; Among them, L Huber Represents the Huber sub-loss of the fundamental phase, η i This represents the Huber sub-loss weight corresponding to the phase map at the i-th scale before unifying the scale. This represents the phase diagram at the i-th scale before scaling. This represents the true phase diagram.

[0036] In this embodiment of the invention, phase maps of three scales are output, and the corresponding phase loss is as follows: ; In the above formula, The diagrams represent the fine-scale phase map, the medium-scale phase map, and the coarse-scale phase map, respectively, with η1~η3 representing the corresponding Huber loss weights. It is understood that loss calculation can only be performed when all phase maps are scaled to the same size as the true phase map. Therefore, in this embodiment of the invention, bilinear interpolation upsampling is used to unify the scale of all phase maps to the scale of the true phase map, with η1~η3 being 1, 0.4, and 0.3 respectively.

[0037] The sine and cosine consistency loss is: ; Where Ω represents the non-zero region in the true phase diagram, and |Ω| represents the total number of elements in the non-zero region. This represents the phase diagram that the system finally recovers. This represents the corresponding true phase diagram; The gradient loss is: ; in, and These represent the calculation of the first-order gradient in the horizontal and vertical directions, respectively. The smoothing loss is: ; It should be noted that the purpose of smoothing loss is to prevent the predicted slope surface from being covered with fine burrs and random noise, and to make the surface smooth. Therefore, smoothing loss does not require the truth value to be involved.

[0038] The edge consistency loss is: ; in, This represents the calculation of the second-order Laplace derivative.

[0039] The multi-scale loss is: ; in, This represents the phase map after downsampling at the j-th mesoscale from the phase map finally recovered by the system. This represents the phase map after downsampling the true phase map at the j-th mesoscale, γ. j This represents the corresponding multi-scale loss weight. In this embodiment of the invention, the phase map finally recovered by the system and its corresponding true phase map are simultaneously downsampled by 1 / 2 and 1 / 4, and the corresponding multi-scale loss is: ; In the above formula, The figures, in sequence, represent the phase map finally recovered by the system, the corresponding phase map after 1 / 2 downsampling, and the corresponding phase map after 1 / 4 downsampling. The numbers represent the true phase map, the true phase map after 1 / 2 downsampling, and the true phase map after 1 / 4 downsampling, respectively, with γ1~γ3 being 1, 0.6, and 0.3, respectively.

[0040] In some embodiments, the basic phase Huber sub-loss is: ; Where Huber represents the circumferential phase Huber loss, which is: ; Where ε represents the phase diagram Compared with the true phase diagram The phase difference between them, δ represents the supervision threshold. In this embodiment of the invention, the phase map is... Compared with the true phase diagram The phase difference between the two phases is wrapped in the interval [-π, π] before the circular phase Huber loss is calculated, thus avoiding the spurious large error caused by the 2π periodicity. The basic phase Huber sub-loss is then expressed as: ; Here, wrap represents the function wrapping.

[0041] In this embodiment of the invention, phase maps of three scales were obtained. The supervision threshold corresponding to the smallest scale phase map was 0.24, the supervision threshold corresponding to the medium scale phase map was 0.20, and the supervision threshold corresponding to the largest scale phase map (i.e. the phase map finally recovered) was 0.18.

[0042] Furthermore, the training hyperparameters in this embodiment of the invention include: updating the network parameters using the AdamW optimizer, with an initial learning rate preferably of 1.5 × 10⁻⁶. 4 The optimal minimum learning rate is 5×10. 7 The preferred weight decay is 8×10. 5 A preheating-cosine annealing learning rate strategy is adopted, with a preferred number of preheating rounds of 5. A gradient pruning strategy is used to limit the gradient norm to within 1.0, and mixed precision training is employed to improve computational efficiency. To further enhance stability during the validation phase, this embodiment introduces an exponential moving average model to smoothly update the main model parameters during training, with a preferred decay coefficient of 0.9992.

[0043] S4: Input the phase-shifting interferometric image sequence to be tested into the phase recovery model obtained in step S3, and output the final phase map.

[0044] To verify the effectiveness of the system and method provided by this invention, comparative experiments were also conducted in this embodiment. Specifically, the test data were grouped and statistically analyzed according to the sample disturbance intensity parameter, dividing the test data into a weak disturbance group, a medium disturbance group, and a strong disturbance group. Using the same strong disturbance group test data, the method and method provided by this invention were compared. Figure 3 (de Groot in the middle), and the weighted polynomial phase-shift demodulation method (corresponding to Figure 3 A comparative experiment was conducted using the Kim / Hibino pair, with root mean square error (RMSE) and mean absolute error (MAE) as evaluation metrics. The RMSE comparison results are as follows: Figure 3 As shown, the MAE comparison results are as follows: Figure 4 As shown. From Figure 3 and Figure 4 As can be seen, this invention achieves the lowest RMSE. This is because the invention adopts a physical prior construction method based on weighted least squares Huber fitting, and combines it with joint optimization strategies such as bidirectional temporal feature extraction and multi-scale deep supervision. This enables the model to fully utilize the physical and temporal information of the interferometric image, effectively suppress the influence of noise and outliers on phase recovery, and improve the accuracy, stability and robustness of phase prediction under complex turbulent conditions.

[0045] This invention also analyzes the changes during the system training process, specifically... Figure 5 The curves showing the change of the loss function for training and validation data during model training are illustrated. Figure 6 The curves showing the change of the root mean square error of the phase on the training and validation data during model training are shown. Figure 7 The curve showing the change in the learning rate during model training is illustrated. Figures 5-7 It can be seen that the training curve of the system and method provided by the present invention decreases steadily and has low RMSE convergence characteristics. This is essentially due to the collaborative constraint mechanism of temporal attention, temporal statistical modeling, multi-scale deep supervision and physical consistency loss in the present invention, which enables the system to gradually approach the true phase solution even without physical initial conditions, thereby achieving a stable, low-error and generalizable phase recovery effect under strong turbulence conditions.

[0046] The method of this invention is applicable not only to simulations or offline datasets, but also to online measurement systems. By simply integrating the image acquisition module with the inference module, rapid phase reconstruction of the sample under test can be achieved. For application scenarios with different resolutions, frame rates, or noise levels, the input size, basic channel count, temporal hidden dimension, loss weights, and training epochs can be adjusted accordingly without departing from the core concept of this invention.

[0047] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0048] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A phase retrieval system for strongly turbulent phase shift interferometry based on spatiotemporal deep networks, characterized in that, include: The frame statistics unit performs statistical calculations on the acquired phase-shifted interferometric image sequence to obtain statistical characteristics; The frame statistics unit calculates the frame mean, frame standard deviation, and frame maximum value of the phase-shifted interferometric image sequence, completes the statistical calculation, and the resulting frame mean, frame standard deviation, and frame maximum value constitute statistical features; The temporal modeling unit extracts temporal information and enhances attention on the input image stack composed of phase-shifted interferometric image sequences to obtain temporal attention features and temporal statistical features. In the temporal modeling unit: the input image stack is expanded in dimension, and then the expanded features are convolved in three dimensions to obtain preliminary features; the preliminary features are convolved in three dimensions and softmax to obtain temporal attention weights, and then the temporal attention weights are added to the preliminary features to obtain temporal attention fusion features. Temporal statistical features are obtained by calculating the temporal mean and temporal maximum response of the temporal attention fusion features respectively. The spatial modeling unit fuses the spatial fusion features composed of phase-shifted interferometric image sequences and statistical features with temporal attention features and temporal statistical features in a spatial direction to obtain spatiotemporal joint features; The phase extraction unit performs step-by-step encoding and decoding operations on the spatiotemporal joint features to obtain multi-scale phase features. In the phase extraction unit, the U-Net network is used to perform step-by-step encoding and decoding operations on the spatiotemporal joint features to obtain multi-scale phase features. The phase retrieval unit recovers phase features at multiple scales, with the recovery of the phase shift feature at the largest scale being the final output phase.

2. The strong turbulent phase shift interferometry phase retrieval system based on spatiotemporal depth networks according to claim 1, characterized in that, In the spatial modeling unit: Spatial statistical features are obtained by performing two-dimensional convolution on spatial fusion features; After channel concatenation and convolution of spatial statistical features, temporal attention features, and temporal statistical features, spatiotemporal joint features are obtained.

3. The strong turbulent phase shift interferometry phase retrieval system based on spatiotemporal depth networks according to claim 1, characterized in that, The phase recovery unit includes an output head corresponding to the phase feature scale; each output head maps the phase feature of the corresponding scale to a two-channel phase representation feature, one channel being a cosine component and the other channel being a sine component; after normalizing the phase representation feature, the phase is obtained through the arctangent function.

4. A phase retrieval method for strong turbulent phase shift interferometry based on spatiotemporal deep networks, characterized in that, include: S1: Preprocess the acquired multiple sets of phase-shifting interferometric image sequences and the corresponding real phase maps to obtain the training set; S2: Construct a strong turbulent phase shift interference phase recovery system based on spatiotemporal depth network as described in any one of claims 1 to 3; S3: Use the training set built in step S1 to train the system built in step S2 to obtain the phase retrieval model; S4: Input the phase-shifting interferometric image sequence to be tested into the phase recovery model obtained in step S3, and output the final phase map.

5. The phase retrieval method for strong turbulent phase shift interference based on spatiotemporal depth networks according to claim 4, characterized in that, In step S3, the system is trained using the following joint loss function: L total =λ1L ph +λ2L sc +λ3L grad +λ4L ms +λ5L smooth +λ6L edge ; Among them, L total L represents the joint loss function. ph L represents phase loss. sc L represents the sine and cosine consistency loss. grad L represents the gradient loss. ms L represents multi-scale loss. smooth L represents the smoothing loss, used to suppress ineffective high-frequency oscillations. edge λ1~λ6 represent the edge consistency loss, and λ1~λ6 represent the loss weights; The phase loss is: ； Among them, L Huber Represents the Huber sub-loss of the fundamental phase, η i This represents the Huber sub-loss weight corresponding to the phase map at the i-th scale before unifying the scale. This represents the phase diagram at the i-th scale before scaling. Represents the true phase diagram; The sine and cosine consistency loss is: ； Where Ω represents the non-zero region in the true phase diagram, and |Ω| represents the total number of elements in the non-zero region. This represents the phase diagram that the system finally recovers. This represents the corresponding true phase diagram; The gradient loss is: ； in, and These represent the calculation of the first-order gradient in the horizontal and vertical directions, respectively. The smoothing loss is: ； The edge consistency loss is: ； in, This represents the calculation of the second-order Laplace derivative; The multi-scale loss is: ； in, This represents the phase map after downsampling at the j-th mesoscale from the phase map finally recovered by the system. This represents the phase map after downsampling the true phase map at the j-th mesoscale, γ. j This represents the corresponding multi-scale loss weight.

6. The phase retrieval method for strong turbulent phase shift interference based on spatiotemporal depth networks according to claim 5, characterized in that, The basic phase Huber sub-loss is: ； Where Huber represents the circumferential phase Huber loss, which is: ； Where ε represents the phase diagram Compared with the true phase diagram The phase difference between them, δ represents the supervision threshold.

Images