Speckle interferometric phase unwrapping method and system based on low-calculation-amount deep learning

Abstract

The invention discloses a speckle interference phase unwrapping method which is suitable for digital speckle interference measurement and realizes low-calculation-amount rapid unwrapping. The method comprises the following steps: firstly, constructing an unwrapping neural network model taking improved ERF-Net as a backbone, and integrally adopting an encoder-decoder-output layer series structure; compressing resolution and extracting deep features at an encoder end through a multi-stage down-sampling module and a non-bottleneck 1D convolution module; a cavity space pyramid pooling (ASPP) module, a non-bottleneck 1D module and a deconvolution layer are fused at a decoder end, and the decoder end is aligned with a corresponding layer of an encoder by using residual connection to recover multi-scale details. A to-be-processed wrapped phase image is input into the model, so that an unwrapped phase result can be obtained in real time; in the model training stage, a loss function based on an L1 norm is adopted for supervision, and the difference between a predicted phase and a real phase is minimized. The method can realize accurate recovery of phase information under the background of high noise and complex interference while keeping the light weight of the network, and is suitable for speckle interference measurement and other scenes requiring high-precision phase data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical measurement and digital image processing, and relates to a method and system for unwrapping speckle interferometric phase patterns by using an efficient residual decomposition network (ERFNet). The method combines multi-scale feature extraction and residual connection of deep learning, introduces an attention mechanism and an atrous spatial pyramid pooling (ASPP), so that the wrapped phase in a high-noise or complex gradient scene can be accurately recovered while keeping a small model size and low computational load. The method can be applied to digital speckle interferometry and other technologies that require high-precision phase information. BACKGROUND

[0002] With the rapid development of deep learning in the fields of computer vision and natural language processing, image processing methods based on convolutional neural networks (CNN) have achieved remarkable results. However, for the problem of complex phase unwrapping, traditional convolutional networks have certain limitations in multi-scale feature processing, computational efficiency, and high-precision reconstruction. In particular, in the phase unwrapping task, due to the influence of speckle noise, high gradient regions, and discontinuous stripes in the speckle interferometric measurement system, the wrapped phase may exhibit sharp jumps or breaks in local regions, making it difficult for the network to maintain good estimates of both details and global phase consistency in the global range. Existing methods mostly use an encoder-decoder network structure, but there is still much room for improvement in processing deep feature transmission, fusing multi-scale context information, and suppressing high noise. Traditional unwrapping algorithms (such as the branch cut method, quality map guided method, least squares method, etc.) are prone to large-area phase breaks or artifacts in the presence of extreme noise, while earlier deep learning methods, although enhanced in model depth and feature expression ability, often come with a large number of parameters and high computational overhead, making them difficult to promote in resource-constrained or real-time processing scenarios. Therefore, how to maintain high-precision phase reconstruction while considering the compactness of the network model and the training and inference efficiency has become a key issue in phase unwrapping algorithm research. The ERFNet-based simultaneous fusion of ASPP and attention mechanism can enhance the network's receptive field while reducing redundant calculations and effectively suppressing noise interference, providing a new approach to high-quality phase unwrapping. SUMMARY

[0003] The present application provides a low-computational-depth learning-based optical speckle interferometric phase unwrapping method and system, which is suitable for unwrapping wrapped phase images in the process of digital speckle interferometry, and solves the technical problems of insufficient anti-interference ability, high model computational complexity, and insufficient multi-scale feature fusion of existing methods in a high-noise environment.

[0004] Step a: constructing an unwrapping neural network model

[0005] A unwrapping model is constructed to improve the ERF-Net as the backbone network, which is composed of an encoder subnetwork, a decoder subnetwork and an output layer connected in turn.

[0006] Optionally, the encoder subnetwork adopts a four-level hierarchical compression structure, the decoder subnetwork realizes phase recovery through multi-scale feature fusion, and the output layer outputs a high-resolution unwrapping phase image through bilinear interpolation.

[0007] Step b: encoder subnetwork design

[0008] In the encoder subnetwork, a plurality of downsampling modules and non-bottleneck one-dimensional convolution modules are arranged.

[0009] Optionally, the downsampling module is composed of a parallel 3x3 depth separable convolution (step size 2) and a 2x2 max pooling layer, and the output is compressed by channel splicing and 1x1 convolution, the resolution is reduced to 1 / 16 of the original image, the calculation amount is reduced by 58%, and the multi-scale semantic information is retained.

[0010] Optionally, the non-bottleneck one-dimensional convolution module is composed of two groups of alternating kx1 and 1xk asymmetric convolution chains, each layer is followed by batch normalization and ReLU activation function, and a dynamic dilated convolution (dilation rate increases exponentially to d=8 according to the level) is introduced, the effective receptive field is expanded to 8 times of the original image, and the parameter amount is reduced by 67% compared with the standard convolution.

[0011] Step c: decoder subnetwork design In the decoder subnetwork, an atrous spatial pyramid pooling (ASPP) module, a non-bottleneck one-dimensional convolution module and a deconvolution layer are integrated, and are interconnected with the encoder subnetwork through residual connection.

[0012] Optionally, the ASPP module includes three parallel 3x3 atrous convolutions (dilation rates are 6, 12 and 18 respectively) and a global average pooling branch, and the output is fused by concatenation and 1x1 convolution, which captures multi-scale context features and reduces the calculation amount by 75%. Optionally, the residual connection splices the upsampling results of each level of the decoder and the corresponding layer features of the encoder, and fuses them through 1x1 convolution to enhance the detail reconstruction capability.

[0013] Step d: phase unwrapping execution

[0014] The unwrapping phase image corresponding to the to-be-processed wrapped phase image is obtained by inputting the wrapped phase image into the neural network model.

[0015] Optionally, the input image is pre-processed by guided filtering (filter radius r = 5, regularization parameter ε = 0.01), and the noise standard deviation is reduced from 30 to 8.7 while retaining the edge gradient feature (error < 0.05 rad / pixel).

[0016] Step e: model training and optimization

[0017] The neural network model is supervised trained by using a loss function based on L1 norm to minimize the difference between the unwrapped phase image and the true phase image.

[0018] Optionally, the L1 loss is defined as the absolute error between the predicted phase and the true phase (weight λ1 = 0.7), and compared with L2 loss, the abnormal value sensitivity is reduced by 42%, and the detail texture reconstruction accuracy is improved by 36%.

[0019] Optionally, Gaussian noise (σ ~ U(0, 30)), salt and pepper noise (density p ~ U(0, 0.1)) and phase jump (jump area ratio 5%-20%) are randomly added to the input phase image in the training stage, to enhance the generalization robustness of the model in high noise scenes.

[0020] 2. A phase unwrapping system, the system comprising:

[0021] An input module: receiving wrapped phase image data output by a digital image acquisition device, supporting real-time data stream input;

[0022] A processing unit: internally integrating the ASPP-ERFNet model trained by the method of claims 1-6, realized by a low-power embedded GPU (such as Jetson TX2) or FPGA, using half-precision floating-point operation (FP16) and deep separable convolution acceleration structure, the model parameter quantity is compressed to 8.69MB, and the operation power consumption is reduced by 65%;

[0023] An evaluation module: using normalized root mean square error (NRMSE), peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) to quantitatively evaluate the unwrapping result;

[0024] A storage module: storing model weight parameters and evaluation results, supporting TB-level data long-term archiving.

[0025] Optionally, the processing unit is configured with an online fine-tuning module based on Adam optimization algorithm, when the evaluation module detects NRMSE > 2.0, the last 100 frames of data are extracted from the cache, and the model is fine-tuned for 10 times with a learning rate of 1e-5, the model switching time is < 1ms, and the model is adapted to different noise scenes. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 The algorithm model of the present application;

[0027] Figure 2 Flowchart of post-training;

[0028] Figure 3(a) is a phase wrapping figure with moderate noise in the test set;

[0029] Figure 3(b) is the corresponding Ground Truth of the phase wrapping figure with moderate noise in the test set;

[0030] Figure 3(c) is an unwrapped phase figure based on Goldstein;

[0031] Figure 3(d) is an unwrapped phase figure based on the unwrapping method of CPULSI;

[0032] Figure 3(e) is an unwrapped phase figure based on the unwrapping method of the present application;

[0033] Figure 3(f) is an unwrapped phase figure based on the unwrapping method of DLPU;

[0034] Figure 3(g) is an unwrapped phase figure based on the unwrapping method of VUR-Net;

[0035] Figure 4(a) is a phase wrapping figure with severe noise in a speckle interference experiment;

[0036] Figure 4(b) is an unwrapped phase figure based on Goldstein;

[0037] Figure 4(c) is an unwrapped phase figure based on the unwrapping method of CPULSI;

[0038] Figure 4(d) is an unwrapped phase figure based on the unwrapping method of the present application;

[0039] Figure 4(e) is an unwrapped phase figure based on the unwrapping method of DLPU;

[0040] Figure 4(f) is an unwrapped phase figure based on the unwrapping method of VUR-Net. DETAILED DESCRIPTION

[0041] The specific implementation steps of the "speckle interference phase unwrapping method and system based on low-computational deep learning" will be described in detail in combination with the practice flow of the present application. It should be understood that the order and details of each step can be changed and expanded according to actual needs, and any replacement or deformation within the scope of the idea and core technical solution of the present application is within the protection scope of the present application.

[0042] Step one: design and construction of network model.

[0043] Figure 1The schematic diagram of the improved ERF-Net network structure of the embodiment of the present disclosure is shown. The network as a whole is composed of an encoder sub-network, a decoder sub-network and an output layer connected in series, which is used to wrap the input phase φ wrap (x, y) is converted to absolute phase φ in real time abs (x, y). The encoder is responsible for compressing spatial resolution and extracting deep semantic features; the decoder is responsible for fusing multi-scale context and gradually restoring resolution; and the output layer produces the final unwrapped result.

[0044] like Figure 1 As shown in Figure 1, the encoder has three downsampling modules at the beginning and end. Taking Layer 1 as an example, the convolution branch first performs a 3×3 convolution with a stride of 2 on the input feature F0, and the number of channels is expanded to C1 after downsampling; the pooling branch then performs maximum pooling with the same stride. The outputs of the two branches are concatenated along the channel dimension and passed through BatchNorm and ReLU to obtain F1. Assuming the convolution kernel size is k=3, padding p=1, and stride s=2, the output size satisfies

[0045]

[0046] Thus, the 256×256 feature map is compressed to 128×128.

[0047] Then as Figure 1 As shown, the non-bottleneck one-dimensional convolution block (NonBottleneck1D) uses two asymmetric convolution kernels of 3×1 and 1×3 in series, as shown in the formula

[0048] F(x)=σ(BN(W 1×3 σ(BN(W 3×1 x)))) (2)

[0049] Transform the feature x and output it through the residual shortcut

[0050] y=x+F(x) (3)

[0051] Effectively alleviate deep layer degradation. Set the hole rate d = 2, 4, 8, 16 for the deep NB1D module cycle, and linearly expand the receptive field to k without increasing the parameters. eff =k+2(d-1).

[0052] The decoder stage is as follows Figure 1 As shown in Figure 1, the Atrous Spatial Pyramid Pooling (ASPP) module is inserted in Layer 17. ASPP consists of five parallel branches: a 1×1 convolution, three 3×3 convolutions with dilation rates of 6, 12, and 18, and a global average pooling branch. Let the output of each branch be z i (i=0,...,4), the splicing result

[0053] Z=Concat(z0,z 1，. ..,z4) (4)

[0054] Compressed to a fixed channel C by 1×1 convolution aspp , and then pass ReLU to get the fusion feature F aspp .

[0055] like Figure 1 As shown, Layer23 uses a transposed convolution with a stride of 2 to upsample the feature map size from 32×32 to 64×64, and combines it with the symmetric encoding layer feature F skip Then the channel attention mechanism ECA-Net is called to calculate the weight of each channel c

[0056] w c =σ(Conv1D k＝3 (GAP(F)) c ) (5)

[0057] And get the recalibrated features This highlights key responses and suppresses redundant information.

[0058] The final output layer is Figure 1 As shown, a 1×1 convolution kernel W is used. 1×1 Decoder results Implementing a linear mapping

[0059]

[0060] Directly generate single-channel absolute phase. If the application scenario requires 2π cycle compensation, it can also output integer cycle counts and Post-processing stage

[0061]

[0062] Get the complete phase.

[0063] Step 2: Construction of phase dataset

[0064] True Phase The generation of consists of two main parts: a linear or nonlinear ramp background and a number of Gaussian spot distributions with different positions, widths, and amplitudes. To generate the initial phase distribution for an image of a specified size, the process first constructs random ramps in the x and y directions using randomly generated gradients m1 and m2, and superimposes a constant bias C to form large-scale phase fluctuations. Then, at a randomly selected center (μ x , μ y ), according to the specified standard deviation σ x and σ yIn the linear superposition of Gaussian function, add multiple Gaussian spots to the phase map to obtain local subtle changes; finally, mix these Gaussian spots into the main slope with a small coefficient to obtain the final real phase distribution I(x, y). In order to further enhance the diversity of the generated image, the gradient and bias of the slope, as well as the number, position, size and amplitude of the Gaussian spots are randomly sampled, so that phase images with different styles are generated in different samples.

[0065] For the generated absolute phase I(x, y), the corresponding wrapped phase can be obtained by means of

[0066] φ(x, y) = wrap(I(x, y)) (8)

[0067] , that is, the core is

[0068] φ(x, y) = ∠(e jI(x，y) ) (9)

[0069] It is worth noting that in actual operation, the generated often needs to be rescaled (such as mapping it to [-4π, 4π]) to ensure that the input range is reasonable during subsequent network training.

[0070] Real speckle interferometry is inevitably affected by noise. In order to simulate this process, noise can usually be added to the complex form of the interference image, and then its phase is extracted. The specific method is as follows:

[0071] (1) Let

[0072] z(x, y) = e jI(x,y) (10)

[0073] Here I(x, y) is the real phase;

[0074] (2) Add independent and identically distributed noise n real and n image following Gaussian distribution to the real and imaginary parts of z respectively, and synthesize noise

[0075] n = n real +j·n image (11)

[0076] Get the complex image with noise

[0077] z noisy (x,y) = z(x, y) + n (12)

[0078] (3) Finally, calculate the wrapped phase

[0079] φ noisy (x,y) = ∠(z noisy(x, y)) (13)

[0080] The noise interferogram is simulated by adding Gaussian noise to the phase image and controlling the signal-to-noise ratio (SNR).

[0081] Step three: loss function design and network training.

[0082] The present application adopts a composite loss function for the phase unwrapping problem, taking into account both local gradient consistency and overall phase accuracy. The following components can be integrated: Residual loss: Minimize the deviation of the wrapped phase and the predicted phase in the gradient layer, and promote accurate recovery of local phase jumps.

[0083]

[0084] L1 loss (Cross Entropy Loss): L1 loss is selected to encourage the closeness of the predicted unwrapped phase image and the real phase image.

[0085]

[0086] Set the initial learning rate (such as 0.001), the optimizer (such as Adam), and the batch size (such as 4), and iterate according to the total training rounds (such as 500 rounds) or the early stopping strategy. Learning rate decay strategy can be used in the later stage of training to make the network more robustly converge. When the validation set error or the comprehensive loss reaches the optimal or tends to be stable, the network weight parameters are saved for subsequent inference and application.

[0087] Step four: model prediction and experimental result analysis.

[0088] Input the unknown wrapped phase image into the trained improved ERFNet model; after the multi-layer residual abstraction of the encoder and the ASPP multi-scale fusion in the decoder, the absolute phase is output; according to the needs, simple post-processing (such as boundary difference correction) can be performed to reduce the micro-error caused by local large jumps, and the final unwrapped result is obtained.

[0089] Embodiment

[0090] The phase wrapping plot of the embodiment 1 of the present application is a medium noise in the test set, as shown in Figure 3(a). Figure 3(b) is the Ground Truth of the corresponding unwrapped phase, Figure 3(c), Figure 3(d), Figure 3(e), Figure 3(f), Figure 3(g) are respectively based on the Goldstein unwrapping method, the CPULSI unwrapping method, the unwrapping method of the present application, the DLPU unwrapping method, the VUR-Net unwrapping. From which it can be seen that our method has the smallest NRMSE, which shows that our algorithm is the best. The phase wrapping plot of the embodiment 1 of the present application is a phase wrapping plot containing serious noise in the speckle interference experiment, as shown in Figure 4(a).

[0091] Figure 4(b), Figure 4(c), Figure 4(d), Figure 4(e), Figure 4(f) are respectively based on the Goldstein unwrapping method, the CPULSI unwrapping method, the unwrapping method of the present application, the DLPU unwrapping method, the VUR-Net unwrapping.

[0092] From the results, it can be seen that our method has the best unwrapping effect.

Claims

1. A low-computation deep learning-based optical speckle interferometry phase unwrapping method, suitable for unwrapping wrapped phase images during digital speckle interferometry, characterized by: The following steps are involved: a) constructing an unwrapping neural network model, wherein the model uses an improved ERF-Net as a backbone network, and the improved ERF-Net is composed of an encoder subnetwork, a decoder subnetwork and an output layer connected in sequence; b) Setting up several downsampling blocks and non-bottleneck 1D convolution blocks in the encoder subnetwork; c) Integrate the Atrous Spatial Pyramid Pooling (ASPP) module, non-bottleneck one-dimensional convolution module, and deconvolution layer into the decoder subnetwork, and interconnect it with the encoder subnetwork through residual connections; d) inputting the wrapped phase image to be processed into the neural network model to obtain a corresponding unwrapped phase image; e) Supervising the neural network model using a loss function based on the L1 norm to minimize the difference between the unwrapped phase image and the true phase image.

2. The method according to claim 1, characterized in that The ASPP module includes at least three parallel convolution branches, each with a dilation rate of 1, 6, 12, or 18, and is fused through cascade and 1×1 convolution after the branch output to obtain contextual features of different receptive fields.

3. The method according to claim 1, characterized in that The non-bottleneck one-dimensional convolution module consists of two sets of alternating one-dimensional convolution layers (k×1 convolution and 1×k convolution) and corresponding dilated convolution layers, and a batch normalization layer and ReLU activation function are set between each convolution layer to expand the effective receptive field while reducing the number of parameters.

4. The method according to claim 1, wherein The downsampling module consists of a 3×3 convolutional layer with a stride of 2 and a 2×2 maximum pooling layer in parallel. The outputs of the convolutional layer and the maximum pooling layer are concatenated in the channel dimension and then subjected to a 1×1 convolution to achieve channel compression, so as to retain more semantic information while reducing the resolution.

5. The method according to claim 1, wherein The loss function based on the L1 norm is an absolute error loss, which can reduce the influence of outliers on gradients compared to the mean square error loss based on the L2 norm, and improve the reconstruction accuracy of detailed textures during training.

6. The method according to claim 1, characterized in that During the training phase, data augmentation is performed on the wrapped phase image. The data augmentation includes randomly adding Gaussian noise, salt and pepper noise, and phase jumps to enhance the generalization robustness of the neural network model in high-noise scenarios.

7. A speckle interferometry phase unwrapping system based on low computational load deep learning, characterized in that: include: a) an input module, configured to receive wrapped phase image data output by a digital image acquisition device; b) a processing unit, electrically connected to the input module, internally integrating an ASPP-ERFNet unwrapping model trained according to any one of the methods of claims 1-6, for performing a phase unwrapping operation on the wrapped phase image data and outputting an unwrapped phase image; c) an evaluation module, communicatively connected to the processing unit, for performing quality evaluation on the unwrapped phase image; d) A storage module, used to store the unpacking model weight parameters and evaluation results.

8. The system according to claim 7, characterized in that The processing unit is implemented by a low-power embedded GPU or FPGA, and utilizes half-precision floating-point operations and a depth-separable convolution acceleration structure to reduce computing power consumption and storage overhead.

9. The system according to claim 7, wherein: The evaluation module uses normalized root mean square error (NRMSE), peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) to quantitatively evaluate the unwrapping results, and sends the evaluation results to the processing unit.

10. The system according to claim 9, characterized in that The processing unit is equipped with an online fine-tuning module based on the Adam optimization algorithm, which incrementally updates the unwrapping model according to the evaluation results to further improve the unwrapping accuracy and adapt to different noise scenarios.

Images