A method for suppressing speckle noise in phase fringe images based on deep learning

Abstract

This invention discloses a method for suppressing speckle noise in phase fringe images based on deep learning, comprising the following steps: S1, collecting a phase fringe image dataset; S2, adding speckle noise of random noise levels to training set data samples; S3, preprocessing the noisy training set data; S4, feeding the preprocessed samples into an asymmetric fusion non-local block structure; S5, inputting the results processed in S4 into an FFDNet network for noise suppression; S6, feeding the denoising results obtained in S5 into an edge extraction module; S7, building and training a network model; S8, adding speckle noise of varying noise levels to test set data samples according to S2; S9, feeding the test set samples obtained in S8 into the network model trained in S7. This deep learning method can suppress speckle noise of varying noise levels without changing network parameters, while preserving image edge detail information.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical detection and image processing, and in particular relates to a method for suppressing speckle noise in phase fringe images based on deep learning. Technical Background

[0002] In recent years, with the development of computers, especially high-resolution charge-coupled devices (CCDs), digital holography and its applications have gained increasing popularity. It has been widely used in fields such as three-dimensional imaging and particle measurement. However, due to the strong coherence of lasers, the optical path can be affected by optical component defects (such as dust, bubbles, scratches, and uneven defects) or turbid media. This can cause speckle particles with randomly fluctuating light intensity in the phase images obtained by digital holography, known as speckle noise. This obscures some of the object's detailed features and, in turn, affects the measurement accuracy of the reconstructed image.

[0003] Traditional speckle noise suppression algorithms include total variation regularization, random resampling masks, three-dimensional block matching (BM3D), Fourier filtering (WFF), and wavelet thresholding. While these filter-based image processing methods suppress speckle noise, they blur image details to varying degrees and struggle to suppress speckle noise at higher levels. Currently used deep learning network denoising algorithms, such as Gan, ResNet, and DNCNN, can only be trained on speckle images with a single noise level and are not suitable for processing speckle images with varying noise levels. Furthermore, they can oversmooth the images, resulting in a loss of detailed image information.

[0004] With the in-depth integration of deep learning and image processing, deep learning-based speckle noise suppression methods have been widely used. In the actual speckle noise suppression process, the causes of speckle noise and the environment vary, resulting in complex speckle noise models and varying noise levels. Therefore, existing deep learning-based speckle noise suppression methods cannot simultaneously suppress speckles of different noise levels. Summary of the Invention

[0005] To address the aforementioned issues, the present invention proposes a deep learning-based method for suppressing speckle noise in phase-fringe images. First, this method preprocesses the dataset, improving the model's generalization capabilities. Second, to obtain more feature information from speckle noise images, spatial pyramid pooling is integrated with a non-local attention mechanism to create an asymmetric fusion non-local block. This captures the connection between two different pixels in the image, ensuring that the output at each position incorporates features from all positions across the entire image, yielding global information. This global information is then fed into an FFDNet network to train speckle noise images with varying noise levels. An edge extraction module is added to the output of the FFDNet network to extract and fuse feature information at different scales to generate an edge map. Finally, the resulting edge map is backpropagated to the FFDNet network to guide the network's noise suppression process, facilitating the recovery of more detailed image features. This method effectively improves speckle noise suppression capabilities across varying noise levels while preserving more edge detail in the image.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for suppressing speckle noise in phase fringe images based on deep learning, comprising the following steps:

[0007] Step S1, collecting a phase fringe image dataset as a dataset sample, and dividing the obtained dataset sample into a training set, a validation set, and a test set in a ratio of 8:1:1;

[0008] Step S2, adding different levels of random speckle noise to the training set obtained by the division in step S1;

[0009] Step S3, preprocessing the training set processed in step S2;

[0010] Step S4, inputting the training set pre-processed in step S3 into the asymmetric fusion non-local block structure to combine the information of the noise image with the global information;

[0011] Step S5, inputting the result obtained by step S4 into the FFDNet network;

[0012] Step S6, inputting the denoising result obtained in step S5 into the edge extraction module, extracting the edge detail features of the phase fringe image, and back-propagating the edge detail features into the FFDNet network, so that the network retains more edge detail features;

[0013] Step S7, building a speckle noise suppression network model for phase fringe images, and using phase fringe images with different noise levels to train the network;

[0014] Step S8, adding speckle noise of different noise levels to the test set obtained by the division in step S1 according to the method described in step S2, and performing normalization processing according to the method described in step S3;

[0015] Step S9: sending the test set obtained in step S8 to the network model trained in step S7 to obtain a speckle noise suppression result.

[0016] Preferably, the step S1 specifically includes:

[0017] Step S101: collect phase fringe image datasets to form the noise-free image data required by the improved FFDNet network model as dataset samples, and the image size is unified to 1024×1024 pixels.

[0018] Step S102, data partitioning, divide the data set samples into three parts: training set, validation set and test set in a ratio of 8:1:1; among them, the training set is used to fit the model to the data samples, and the training error is gradient descent and weight parameter training is performed during the training process; the validation set is used to adjust the model's hyperparameters and preliminarily evaluate the model's capabilities; the test set is used to evaluate the generalization ability of the final model.

[0019] Preferably, the step S2 specifically includes:

[0020] The phase fringe images in the training set are added with speckle noise that obeys the multiplicative component of Rayleigh distribution and the additive component of Gaussian distribution as the noise model required for the training model. The noise model can be expressed as:

[0021] n=R(x)+N(0,δ 2 ) (1)

[0022] In formula (1), n is the image after adding speckle noise; x is the noise-free image; R(x) is the Rayleigh distribution with size parameter; N(0,δ 2 ) has an expected value of 0 and a variance of δ 2 Gaussian distribution; different standard deviations δ determine different levels of noise, and the noise level is set to [0,75], and finally a noise image with a noise level of [0,75] corresponding to the original phase fringe image is formed.

[0023] Preferably, the step S3 specifically includes:

[0024] Step S301: perform a linear function normalization operation on each piece of data. The formula is:

[0025]

[0026] In formula (2), p' is the normalized pixel value; p is the original pixel value; min(p) is the minimum pixel value in the sample data; and max(p) is the maximum pixel value in the sample data.

[0027] In step S302, the training image is randomly cropped into image blocks of 160×160 pixels and rotated for data augmentation. This can expand the image data, enhance the model stability, and improve the computational efficiency of the network. The data augmentation method is as follows:

[0028] Rotate 90: Rotate the image block 90° counterclockwise;

[0029] Rotate 180: Rotate the image block 180° counterclockwise;

[0030] Rotate 270: Rotate the image block 270° counterclockwise;

[0031] Flip: flip the image block up and down along the x-axis;

[0032] Flip (rotate 90 degrees): rotate the image block 90 degrees counterclockwise, and then flip it upside down;

[0033] Flip (rotate 180): rotate the image block 180° counterclockwise and then flip it upside down;

[0034] Flip (rotate 270): Rotate the image block 270° counterclockwise and then flip it upside down.

[0035] Step S303 , forming a training image pair with the noise-free phase fringe image block processed in step S302 and the phase fringe image block containing speckle noise at the corresponding position.

[0036] Preferably, the step S4 specifically includes:

[0037] In step S401, an asymmetric fusion non-local block structure is introduced at the front end of the FFDNet network. The asymmetric fusion non-local block is used to count global information, capture the connection between two different pixels on the image, and calculate the response of each pixel position in the image as the weighted sum of the features of all positions except the pixel. The response value is passed to FFDNet to enhance the denoising effect of the network.

[0038] In step S402 , a spatial pyramid pooling operation is added to the asymmetric fusion non-local block structure to provide sufficient feature statistical information about global scene semantic clues, thereby reducing the amount of computation without causing a decrease in the performance of the non-local blocks.

[0039] Preferably, the step S5 specifically includes:

[0040] In step S501, the input noisy image is downsampled using an inverse pixel reconstruction method to reconstruct four downsampled subimages. The size of the noisy global information image is C × W × H, and the size of the downsampled subimages is 4C × W / 2 × H / 2. Downsampling subimages effectively increases the network's receptive field, improves network convolution efficiency, and reduces memory usage.

[0041] At the same time, the downsampled sub-image is concatenated with an image with adjustable noise level as the input of the FFDNet network with a size of (4C+1)×W / 2×H / 2.

[0042] In step S502, the FFDNet network depth is set to 15 layers. The first layer consists of a convolutional layer and a rectified linear unit, that is, the input image is convolved and then activated using the rectified linear unit; the second to thirteenth layers use a combination of convolutional layers, rectified linear units, and batch normalization layers, that is, a batch normalization layer is added between the convolutional layer and the rectified linear unit; the fifteenth layer only uses image convolution to reconstruct the output layer, and the convolution kernel size is 3×3 pixels.

[0043] In step S503 , an upsampling operation is performed by a pixel reassembly method to restore the noise residual image to its original size. The size of the noise residual image is (4C+1)×W / 2×H / 2, and the size of the output image after the upsampling operation is C×W×H.

[0044] Preferably, the step S6 specifically includes:

[0045] In step S601, the edge extraction network has a total of 7 convolutional layers with a convolution kernel size of 3×3 pixels. A 2×2 pixel maximum pooling layer is connected after the second and fourth layers respectively, and the width and height of the input size of the third and fifth convolution layers are changed to half of the previous layer.

[0046] In step S602, after the second, fourth, and seventh convolution layers, the side outputs are extracted and upsampled using bilinear interpolation to align the side output image size with the input image. Finally, the side outputs of different scales are fused pixel by pixel to generate an edge map, which is then used to guide the network's denoising process via backpropagation.

[0047] Preferably, the step S7 specifically includes:

[0048] Step S701: Global information of training set data obtained by asymmetric fusion of non-local block structure statistics The noise level map M is connected in series and used as the input of the FFDNet network; the noise image As a label to train the FFDNet network, the network output is the noise residual map obtained by residual learning Using global information images containing noise Noisy residual map after subtracting the network predictions Get the denoised target image The denoised target image is input into the edge extraction module to guide the network to generate a noise-suppressed image y with more edge detail information and texture features;

[0049] Step S702: Set the model training parameters. The model optimization method is adaptive moment estimation. The initial learning rate is 0.001, the learning rate adjustment method is multi-step decay, and the batch size is 64. The cost function of the network is the mean absolute value error between the noise residual map estimated by the network and the noise map:

[0050]

[0051] In formula (3), is the jth input noise image; m is the total number of training images; To correspond Noise residual map of M j To correspond Noise level diagram; θ is the network parameter; N j To correspond Noise diagram; j is 1.....m.

[0052] Preferably, the step S8 specifically includes:

[0053] Speckle noise with three different noise levels of 25, 50 and 75 is added to the phase fringe images in the test set and normalized.

[0054] Preferably, the step S9 specifically includes:

[0055] The test set images are input into the trained model to obtain the denoising results, and the four evaluation parameters of peak signal-to-noise ratio, structural similarity, equivalent view number, and speckle suppression index are calculated.

[0056] Step S901: The peak signal-to-noise ratio (PSNR) is expressed as:

[0057]

[0058]

[0059] In formula (4), MSE is the mean square error between the original image and the denoised image; in formula (5), x is the noise-free image; y is the image after noise suppression; mn is the total number of pixels in the training image of size m×n; i, j are the pixel values in the i-th row and j-th column.

[0060] Step S902: The mathematical expression of structural similarity SSIM is:

[0061]

[0062] In formula (6), u and v are two local windows of size W×W in the noise-free image and the noise-suppressed image respectively; μ u and μ v are the average values of pixel grayscale in the two windows respectively; and are the variances of the pixel grayscale in the two windows; uv is the covariance of the two windows u and v; constants C1 = 2.55, C2 = 7.65.

[0063] Step S903: The mathematical expression of the equivalent view number ENL is:

[0064]

[0065] In formula (7), μ y is the average value of the image after noise suppression; δ y is the standard deviation of the image after noise suppression.

[0066] In step S904, the mathematical expression of the speckle suppression index SSI is:

[0067]

[0068] In formula (8), δ y is the standard deviation of the image after noise suppression; δ n is the standard deviation of the noisy image; μ y is the mean value of the image after noise suppression; μ n is the mean value of the noisy image.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] In the actual speckle noise suppression process, the noise model is complex and the noise level varies. Because traditional algorithms and conventional networks target specific noise levels, it is difficult to simultaneously suppress speckle noise of different noise levels. Furthermore, image edge detail information is severely lost, which reduces the quality of digital holographic image reconstruction and limits the application of noise suppression methods. Compared with traditional algorithms and conventional network models, this paper proposes a new network structure to effectively suppress speckle noise in phase fringe images. Based on the FFDNet network, this method adds an asymmetric fusion non-local block structure to extract the connection between two pixels in the random noise level image, which overcomes the shortcomings of conventional network models in suppressing noise of different levels and significantly reduces memory usage. In addition, the network adds an edge extraction module at the end of the FFDNet network. This module generates an edge map by pixel-by-pixel summing of side outputs at different scales. Backpropagation guides the network denoising process, ensuring that the network output better preserves image edge detail information. Therefore, the proposed processing strategy has stronger denoising and edge restoration capabilities, significantly suppresses speckle noise of different noise levels, and preserves more image edge detail information, not only improving measurement accuracy but also reducing the burden on hardware equipment.

[0071] The proposed speckle noise suppression method facilitates the neural network to extract deep features of noisy images through asymmetric fusion of non-local blocks (AFNB). It is suitable for speckle noise suppression with different noise levels and improves the noise suppression effect. At the same time, the added edge extraction module (VEM) retains more image edge detail features. This method not only expands the scope of noise suppression, but also improves the ability to extract image edge details. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 This is a flow chart of a deep learning speckle noise suppression method for phase fringe images according to an embodiment of the present invention;

[0073] Figure 2 Phase fringe images containing speckle noise of different noise levels according to an embodiment of the present invention;

[0074] Figure 3 The network model of the embodiment of the present invention is Figure 2 Speckle noise suppression results of phase fringe images. DETAILED DESCRIPTION

[0075] The invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0076] Reference Figure 1 The present invention discloses a method for suppressing speckle noise in phase fringe images based on deep learning. The specific implementation steps are as follows:

[0077] Step S101: collect phase fringe image datasets to form the noise-free image data required for training the improved FFDNet network model. As dataset samples, the image size is unified to 1024×1024 pixels to facilitate model reading of data.

[0078] Step S102: To improve the generalization ability of the model, the dataset samples are divided into three parts: training set, validation set, and test set in a ratio of 8:1:1. In the process of selecting the training set and the test set, it is ensured that the selected training set samples and the test set samples cover phase fringe distributions of various sparsity levels.

[0079] Reference Figure 2 Step S2 of the present invention includes: using the Numpy extension package in Python to add speckle noise that obeys the multiplicative component of Rayleigh distribution and the additive component of Gaussian distribution to the phase fringe image in the training set as the noise image required for the training model; the noise model can be expressed as:

[0080] n=R(x)+N(0,δ 2 ) (1)

[0081] In formula (1), n is the image after adding speckle noise; x is the noise-free image; R(x) is the Rayleigh distribution with size parameter; N(0,δ 2 ) has an expected value of 0 and a variance of δ 2 The Gaussian distribution of [0, 75] is used; different standard deviations δ determine different levels of noise. Setting the noise level to [0, 75] ultimately produces a noise image with a noise level of [0, 75] corresponding to the original phase fringe image. The encoding format is UTF-8, and the noise image label is saved in BMP format.

[0082] Step S301: Data normalization performs Min-Max Scaling on each piece of data. The formula is:

[0083]

[0084] In formula (2), p' is the normalized pixel value; p is the original pixel value; min(p) is the minimum pixel value in the sample data; and max(p) is the maximum pixel value in the sample data.

[0085] In step S302, the training image is randomly cropped into image blocks of 160×160 pixels and rotated for data augmentation. This can expand the image data and enhance the model stability while improving the computational efficiency of the network. The data augmentation method is as follows:

[0086] Rotate 90: Rotate the image block 90° counterclockwise;

[0087] Rotate 180: Rotate the image block 180° counterclockwise;

[0088] Rotate 270: Rotate the image block 270° counterclockwise;

[0089] Flip: flip the image block up and down along the x-axis;

[0090] Flip (rotate 90 degrees): rotate the image block 90 degrees counterclockwise, and then flip it upside down;

[0091] Flip (rotate 180): rotate the image block 180° counterclockwise and then flip it upside down;

[0092] Flip (rotate 270): Rotate the image block 270° counterclockwise and then flip it upside down.

[0093] Step S303 , forming a training image pair with the noise-free phase fringe image block processed in step S302 and the phase fringe image block containing speckle noise at the corresponding position.

[0094] Step S401: Introduce an asymmetric fusion non-local block structure in the front end of the FFDNet network. First, the input image X∈R is transformed by three 1×1 convolutions. C×H×W Transformed into three feature matrices θ∈R C×H×W θ,γ∈R C×H×W ,Will Feature conversion to Secondly, use SPP to downsample the feature matrices θ and γ to obtain a smaller output, θ P ∈R C×S , γ P ∈R S×C . Again, the feature matrix and θ P Multiply and normalize by SoftMax function to get V∈R N×S Then, the feature matrix V is combined with γ P Multiply to get O∈R N×C Finally, the new output features are concatenated with the original input features after a 1×1 convolution transformation: This is the final output of the module.

[0095] In step S402, a spatial pyramid pooling operation is added to the asymmetric fusion non-local block structure. Spatial pyramid pooling performs maximum pooling operations on θ and γ at different scales to improve the network's receptive field, effectively avoiding image distortion caused by cropping and scaling operations on image regions, and also solves the problem of repeated feature extraction related to the image, saving computational costs. The convolution kernel size selected for spatial pyramid pooling is {1, 3, 6, 8}, and the number of sampling anchor points is S = ∑ n∈{1,3,6,8} n2 =110, and we get the downsampled θ P ∈R C×S , γ P ∈R S×C The network's input data channel is 1, and the image width and height are both 1024. Therefore, asymmetric matrix multiplication can save 1024 × 1024 / 110 ≈ 9532 times the computation time. In addition, spatial pyramid pooling provides sufficient feature statistics about global scene semantic cues, reducing the computational effort without degrading the performance of non-local operation blocks.

[0096] In step S501, the input noisy image is downsampled using an inverse pixel reconstruction method to reconstruct four downsampled subimages. The input noisy global information image is of size C × W × H, and the downsampling factor is 2. The resulting downsampled subimages are of size 4C × W / 2 × H / 2. Downsampling subimages can effectively increase the network's receptive field, improve network convolution efficiency, accelerate training and testing, and reduce memory usage.

[0097] At the same time, the downsampled sub-image is concatenated with an image with an adjustable noise level as the input of the FFDNet network with a size of (4C+1)×W / 2×H / 2.

[0098] In step S502, the FFDNet network depth is set to 15 layers. The first layer consists of a convolutional layer and a rectified linear unit, that is, the input image is convolved and then activated using a rectified linear unit; the second to thirteenth layers use a combination of convolutional layers, rectified linear units, and batch normalization layers; the fifteenth layer only uses convolution to reconstruct the output layer, and the convolution kernel size is 3×3 pixels.

[0099] In step S503, an upsampling operation is performed using a pixel reshaping method to restore the size of the noisy residual image to its original size. The size of the noisy residual image is (4C+1)×W / 2×H / 2, the upsampling factor is 2, and the size of the output image is C×W×H. The pixel reshaping method uses sub-pixel convolution to expand the feature map to obtain a high-resolution image, solving some artifacts caused by interpolation and transposed convolution, and also solving the checkerboard effect.

[0100] In step S601, the edge extraction network has a total of 7 convolutional layers, and the convolution kernel size is 3×3 pixels. A 2×2 pixel maximum pooling layer is connected after the second and fourth layers respectively, and the width and height of the input size of the third and fifth convolution layers are changed to half of the previous layer.

[0101] In step S602, after the second, fourth, and seventh convolution layers, the side outputs are extracted and upsampled using bilinear interpolation to align the side output image size with the input image. Finally, the side outputs of different scales are fused pixel by pixel through a fusion process to generate an edge map, and backpropagation is used to guide the network's denoising process.

[0102] Step S701: Asymmetric fusion of the global information of the training set data obtained by non-local block structure statistics Downsample and concatenate the noise level map M as the input of the FFDNet network.

[0103] Step S702: The global information image containing noise Subtract the corresponding pixels from the noise-free image x to obtain the noise image As the label training FFDNet network, the network output is the noise residual map obtained by residual learning

[0104] Step S703: using the global information image containing noise Noisy residual map after subtracting the network predictions Get the denoised target image

[0105] Step S704: Input the denoised target image into the edge extraction module, and obtain the edge map by fusing the side outputs of different scales. Guide the network to generate a noise-suppressed image y with more edge detail information and texture features.

[0106] Step S705: Set the model training parameters: the model optimization method is adaptive moment estimation, the initial learning rate is 0.001, the learning rate adjustment method is multi-step decay, the batch size is 64, and the hyperparameter λ is used to control the rate of loss decrease. In this method, λ is set to 0.5. The cost function of the network is the mean absolute error between the noise residual map estimated by the network and the noise map:

[0107]

[0108] In formula (3), is the jth input noise image; m is the total number of training images; To correspond Noise residual map of M j To correspond Noise level diagram; θ is the network parameter; N j To correspond Noise diagram; j is 1.....m.

[0109] Step S8: adding speckle noise of three different noise levels of 25, 50, and 75 to the phase fringe images in the test set, and performing normalization processing.

[0110] Reference Figure 3 , step S9 of the present invention includes: inputting the test set image into the trained model to obtain the denoising result, and calculating its four evaluation parameters of peak signal-to-noise ratio, structural similarity, equivalent view number and speckle suppression index.

[0111] Step S901: The peak signal-to-noise ratio (PSNR) is expressed as:

[0112]

[0113]

[0114] In formula (4), MSE is the mean square error between the original image and the denoised image; in formula (5), x is the noise-free image; y is the image after noise suppression; mn is the total number of pixels in the training image of size m×n; i, j are the pixel values in the i-th row and j-th column.

[0115] Step S902: The mathematical expression of structural similarity SSIM is:

[0116]

[0117] In formula (6), u and v are two local windows of size W×W in the noise-free image and the noise-suppressed image respectively; μ u and μ v are the average values of pixel grayscale in the two windows respectively; and are the variances of the pixel grayscale in the two windows; uv is the covariance of the two windows u and v; constants C1 = 2.55, C2 = 7.65. As shown in Table 1: Figure 3 PSNR, SSIM, SSI, and ENL values corresponding to the phase fringe image after speckle noise suppression.

[0118] Table 1

[0119]

[0120] Step S903: The mathematical expression of the equivalent view number ENL is:

[0121]

[0122] In formula (7), μ y is the average value of the denoised image; δ y is the standard deviation of the image after noise suppression.

[0123] In step S904, the mathematical expression of the speckle suppression index SSI is:

[0124]

[0125] In formula (8), δ y is the standard deviation of the image after noise suppression; δ n is the standard deviation of the noisy image; μ y is the mean value of the image after noise suppression; μ n is the mean value of the noisy image.

[0126] After the above nine steps, the phase fringe image speckle noise suppression model based on deep learning proposed in the present invention can be obtained.

[0127] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection described in the claims.

Claims

1. A method for suppressing speckle noise in phase fringe images based on deep learning, characterized in that: The following steps are involved: Step S1, collecting a phase fringe image dataset as a dataset sample, and dividing the obtained dataset sample into a training set, a validation set, and a test set in a ratio of 8:1:1; Step S2, adding different levels of random speckle noise to the training set obtained by the division in step S1; Step S3, preprocessing the training set processed in step S2; Step S4, inputting the training set pre-processed in step S3 into the asymmetric fusion non-local block structure to combine the information of the noise image with the global information; Step S5, inputting the result obtained by step S4 into the FFDNet network; Step S6, inputting the denoising result obtained in step S5 into the edge extraction network, extracting edge detail features of the phase fringe image, and back-propagating the edge detail features into the FFDNet network, so that the network retains more edge detail features; Step S7, building a speckle noise suppression network model for phase fringe images, and using phase fringe images with different noise levels to train the network; Step S8, adding random speckle noise of different noise levels to the test set obtained by the division in step S1 according to the method described in step S2, and performing normalization processing according to the method described in step S3; Step S9, feeding the test set obtained in step S8 into the network model trained in step S7 to obtain a speckle noise suppression result; The step S4 specifically includes: Step S401: Introduce an asymmetric fusion non-local block structure at the front end of the FFDNet network. The asymmetric fusion non-local block is used to count global information, capture the connection between two different pixels on the image, calculate the response of each pixel position in the image as the weighted sum of the features of all positions except the pixel, and pass the response value to FFDNet to enhance the denoising effect of the network; In step S402 , a spatial pyramid pooling operation is added to the asymmetric fusion non-local block structure to provide sufficient feature statistical information about global scene semantic clues, thereby reducing the amount of computation without causing a decrease in the performance of the non-local blocks.

2. The method for suppressing speckle noise in phase fringe images based on deep learning according to claim 1, characterized in that: The step S1 specifically includes: Step S101, collecting a phase fringe image dataset to form the noise-free image data required by the improved FFDNet network model as a dataset sample, and the image size is unified to 1024×1024 pixels; Step S102, data partitioning, divide the data set samples into three parts: training set, validation set and test set in a ratio of 8:1:1; among them, the training set is used to fit the model to the data samples, and the training error is gradient descent and weight parameter training is performed during the training process; the validation set is used to adjust the model's hyperparameters and preliminarily evaluate the model's capabilities; the test set is used to evaluate the generalization ability of the final model.

3. The method for suppressing speckle noise in phase fringe images based on deep learning according to claim 1, characterized in that: The step S2 specifically includes: The phase fringe images in the training set are added with speckle noise that obeys the multiplicative component of Rayleigh distribution and the additive component of Gaussian distribution as the noise model required for the training model. The noise model can be expressed as: n=R(x)+N(0,δ 2 ) (1) In formula (1), n is the image after adding speckle noise; x is the noise-free image; R(x) is the Rayleigh distribution with size parameter; N(0,δ 2 ) has an expected value of 0 and a variance of δ 2 Gaussian distribution; different standard deviations δ determine different levels of noise, and the noise level is set to [0,75], and finally a noise image with a noise level of [0,75] corresponding to the original phase fringe image is formed.

4. The method for suppressing speckle noise in phase fringe images based on deep learning according to claim 1, characterized in that: The step S3 specifically includes: Step S301: perform a linear function normalization operation on each piece of data. The formula is: In formula (2), p' is the normalized pixel value; p is the original pixel value; min(p) is the minimum pixel value in the sample data; max(p) is the maximum pixel value in the sample data; In step S302, the training image is randomly cropped into image blocks of 160×160 pixels and rotated for data augmentation. This can expand the image data, enhance the model stability, and improve the computational efficiency of the network. The data augmentation method is as follows: Rotate 90: Rotate the image block 90° counterclockwise; Rotate 180: Rotate the image block 180° counterclockwise; Rotate 270: Rotate the image block 270° counterclockwise; Flip: flip the image block up and down along the x-axis; Flip: Rotate the image block 90° counterclockwise and then flip it upside down; Flip: Rotate the image block 180° counterclockwise and then flip it upside down; Flip: Rotate the image block 270° counterclockwise and then flip it upside down; Step S303 , forming a training image pair with the noise-free phase fringe image block processed in step S302 and the phase fringe image block containing speckle noise at the corresponding position.

5. The method for suppressing speckle noise in phase fringe images based on deep learning according to claim 1, characterized in that: The step S5 specifically includes: Step S501: The input noisy image is downsampled using a pixel inverse reconstruction method to reconstruct four downsampled sub-images. The size of the noisy global information image is C×W×H, and the size of the downsampled sub-images is 4C×W / 2×H / 2. The downsampled sub-images can effectively increase the network receptive field, improve the network convolution efficiency, and reduce the memory burden. At the same time, the downsampled sub-image is concatenated with an image with adjustable noise level as the input of the FFDNet network with a size of (4C+1)×W / 2×H / 2; In step S502, the FFDNet network depth is set to 15 layers. The first layer consists of a convolutional layer and a rectified linear unit, that is, the input image is convolved and then activated using the rectified linear unit; the second to thirteenth layers use a combination of convolutional layers, rectified linear units, and batch normalization layers, that is, a batch normalization layer is added between the convolutional layers and the rectified linear units; the fifteenth layer uses only image convolution to reconstruct the output layer, and the convolution kernel size is 3×3 pixels; In step S503 , an upsampling operation is performed by a pixel reassembly method to restore the noise residual image to its original size. The size of the noise residual image is (4C+1)×W / 2×H / 2, and the size of the output image after the upsampling operation is C×W×H.

6. The method for suppressing speckle noise in phase fringe images based on deep learning according to claim 1, characterized in that: The step S6 specifically includes: Step S601: The edge extraction network has a total of 7 convolutional layers with a convolution kernel size of 3×3 pixels. A 2×2 pixel maximum pooling layer is connected after the second and fourth layers respectively, and the width and height of the input size of the third and fifth convolution layers are reduced to half of the previous layer. In step S602, after the second, fourth, and seventh convolution layers, the side outputs are extracted respectively, and up-sampled by bilinear interpolation to make the side output image size consistent with the input image size; finally, the side outputs of different scales are fused by pixel-by-pixel addition to obtain an edge map, and the network denoising process is guided by back propagation.

7. The method for suppressing speckle noise in phase fringe images based on deep learning according to claim 1, characterized in that: The step S7 specifically includes: Step S701: Global information of training set data obtained by asymmetric fusion of non-local block structure statistics The noise level map M is connected in series and used as the input of the FFDNet network; the noise image As a label to train the FFDNet network, the network output is the noise residual map obtained by residual learning Using global information images containing noise Noisy residual map after subtracting the network predictions Get the denoised target image The denoised target image is input into the edge extraction network to guide the network to generate a noise-suppressed image y with more edge detail information and texture features; Step S702: Set the model training parameters. The model optimization method is adaptive moment estimation. The initial learning rate is 0.001, the learning rate adjustment method is multi-step decay, and the batch size is 64. The cost function of the network is the mean absolute value error between the noise residual map estimated by the network and the noise map: In formula (3), is the jth input noise image; m is the total number of training images; To correspond Noise residual map of M j To correspond Noise level diagram; θ is the network parameter; N j To correspond The noise map of ; j is a positive integer from 1 to m.

8. The method for suppressing speckle noise in phase fringe images based on deep learning according to claim 1, characterized in that: The step S8 specifically includes: Speckle noise with three different noise levels of 25, 50 and 75 is added to the phase fringe images in the test set and normalized.

9. The method for suppressing speckle noise in phase fringe images based on deep learning according to claim 1, characterized in that: The step S9 specifically includes: The test set images are input into the trained speckle noise suppression network model to obtain the denoising results, and the four evaluation parameters of peak signal-to-noise ratio, structural similarity, equivalent view count, and speckle suppression index are calculated; Step S901: The peak signal-to-noise ratio (PSNR) is expressed as: In formula (4), MSE is the mean square error between the original image and the denoised image; in formula (5), x is the noise-free image; y is the image after noise suppression; mn is the total number of pixels in the training image of size m×n; i, j are the pixel values in the i-th row and j-th column; Step S902: The mathematical expression of structural similarity SSIM is: In formula (6), u and v are two local windows of size W×W in the noise-free image and the noise-suppressed image respectively; μ u and μ v are the average values of pixel grayscale in the two windows respectively; and are the variances of the pixel grayscale in the two windows; uv is the covariance of the two windows u and v; constants C1 = 2.55, C2 = 7.65; Step S903: The mathematical expression of the equivalent view number ENL is: In formula (7), μ y is the average value of the image after noise suppression; δ y is the standard deviation of the image after noise suppression; In step S904, the mathematical expression of the speckle suppression index SSI is: In formula (8), δ y is the standard deviation of the image after noise suppression; δ n is the standard deviation of the noisy image; μ y is the average value of the image after noise suppression; μ n is the average value of the noisy image.

Images