Particle three-view image extraction method based on three-color mask PIV technology

Abstract

The invention provides a particle three-view image extraction method based on a three-color mask PIV (particle image velocimetry) technology. The method comprises the following steps: acquiring an original mosaic image through a three-color mask single-color camera; constructing a neural network model composed of an image feature extraction module and an image reconstruction module; decomposing the original mosaic image into four channels; the four channels respectively represent different color information; gaussian blur operation is carried out on the four channels to obtain a preprocessed image; performing first feature extraction on the preprocessed image through an initial convolutional layer built in an image feature extraction module to obtain a preliminary image; performing multiple convolution and activation operations on the initial image through an activation function to obtain a feature image; and performing image reconstruction on the feature image through an image reconstruction module to obtain a three-view image. According to the method, the feature extraction efficiency is improved, and the integrity, accuracy and quality of the three-view image are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a particle three-viewing angle image extraction method based on a three-color mask PIV technology. Background Art

[0002] The three-color mask PIV technology enables a single camera to perform 3D-3C measurement of the velocity field. The three-color mask PIV uses a three-color mask aperture to modulate the imaging optical path of the color camera, combines color information with perspective information, and enables the RGB three channels of the photosensitive chip to record the three perspective images of the tracer particles in full frame and full resolution, which can effectively reduce the requirements for camera resolution and improve the camera sampling frequency. In this process, high-quality particle three-perspective image extraction is crucial to improving the accuracy of velocity field measurement. Ideally, each pixel of the color camera should be equipped with three image sensors to record the information transmitted by the three bands of red, green, and blue. However, the high manufacturing cost and complex design limit the popularity and application scope of three-sensor color cameras. At present, most color cameras use a single image sensor to complete the recording of color information. The single-sensor camera collects image information by covering the color filter array (CFA) on the image sensor. The CFA array is composed of a group of spectral arrays, and only one color component is allowed to pass through each pixel position. In this way, the value of each pixel can only record one color component, and the other two color components are missing, thereby generating an original mosaic image. Currently, the methods for demosaicing original mosaic images include the nearest neighbor method: interpolating missing pixel values ​​by using the sampling value of the nearest pixel point; the demosaicing method based on deep learning can extract image features from a large number of data sets, and find the mapping relationship between the original image and the mosaic image through continuous iteration, so as to obtain better demosaicing effect.

[0003] However, the High Quality Linear Interpolation (HQLI) algorithm in the nearest neighbor method does not consider the edge direction of the image, which may lead to some errors in the detail texture and high-frequency areas of the image. The de-mosaicing method based on deep learning still makes the running time not advantageous due to the large number of cross-layer connections in the network, and even takes longer than other networks with more parameters, and is not suitable for practical application in edge computing. Therefore, it is very necessary to design a particle three-view image extraction method based on three-color mask PIV technology. Summary of the invention

[0004] The purpose of the present invention is to provide a particle three-view image extraction method based on three-color mask PIV technology to improve the efficiency of feature extraction and improve the integrity, accuracy and quality of three-view images.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A particle three-view image extraction method based on three-color mask PIV technology includes the following steps:

[0007] The original mosaic image is obtained by a three-color mask single color camera;

[0008] Construct a neural network model consisting of an image feature extraction module and an image reconstruction module; the image feature extraction module is connected to the image reconstruction module; the neural network model is constructed according to the U-Net++ structure;

[0009] The original mosaic image is processed by the image feature extraction module to obtain a feature image. The specific steps of processing the original mosaic image by the image feature extraction module include:

[0010] Decompose the original mosaic image into four channels; the four channels represent different color information respectively;

[0011] Perform Gaussian blur operation on the four channels respectively to obtain the preprocessed image;

[0012] Performing a first feature extraction on the preprocessed image through the initial convolution layer built into the image feature extraction module to obtain a preliminary image;

[0013] Perform multiple convolution and activation operations on the preliminary image through the activation function to obtain the feature image;

[0014] The feature image is reconstructed through the image reconstruction module to obtain a three-view image.

[0015] Optionally, the image feature extraction module includes a plurality of Gaussian smoothing units and a plurality of feature extraction units; the Gaussian smoothing units are connected to the feature extraction units, and the feature extraction units are connected to the image reconstruction module;

[0016] The Gaussian smoothing unit is used to perform Gaussian blur processing on the original mosaic image; the feature extraction unit is used to extract the features of the original mosaic image.

[0017] Optionally, the image reconstruction module includes: a plurality of splicing units, a plurality of reconstruction units and a plurality of upsampling layers; the splicing units are respectively connected to the feature extraction unit and the reconstruction unit, and the reconstruction unit is connected to the upsampling layer;

[0018] The splicing unit is used to splice the feature images; the reconstruction unit is used to perform convolution reconstruction operations on the feature images; and the upsampling layer is used to restore the resolution of the feature images.

[0019] Optionally, the feature image is reconstructed by an image reconstruction module to obtain a three-view image, and the specific steps include:

[0020] Perform a splicing operation on the feature images to obtain a spliced ​​image;

[0021] The spliced ​​images are integrated and feature enhanced to obtain an integrated image;

[0022] The integrated image is upsampled to obtain a three-view image.

[0023] Optionally, constructing a neural network model consisting of an image feature extraction module and an image reconstruction module also includes:

[0024] The neural network model is trained through the final loss function; the final loss function includes: structural similarity loss function and mean square error loss function; the expression of the final loss function is: Among them, L is the final loss function, is the structural similarity loss function value of the first output layer of the image reconstruction module, is the structural similarity loss function value of the second output layer of the image reconstruction module, is the structural similarity loss function value of the third output layer of the image reconstruction module, is the mean square error loss function value of the first output layer of the image reconstruction module, is the mean square error loss function value of the second output layer of the image reconstruction module, is the mean square error loss function value of the third output layer of the image reconstruction module, epoch is the number of iterations of the neural network model, SSIM is the structural similarity loss function, and MSE is the mean square error loss function.

[0025] Optionally, the neural network model is trained using the final loss function, and the specific steps include:

[0026] The neural network model is iterated for the first time using the structural similarity loss function to obtain a first iteration model; the number of iterations of the first iteration is 10;

[0027] The first iteration model is iterated for the second time using the mean square error loss function to obtain a trained neural network model.

[0028] Optionally, the calculation formula of the structural similarity loss function is: Among them, x and y are the corresponding blocks of the image, μ x and μy are the average values ​​of x and y, respectively, x , σ y are the variances of x and y, σ xy is the covariance between x and y, C1 and C2 are constants, and SSIM is the structural similarity loss function.

[0029] Optionally, the mean square error loss function is calculated as: Wherein, I and K are images to be compared, m and n are the width and height of the image respectively, and I(i, j) and K(i, j) are the pixel values ​​at position (i, j).

[0030] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the particle three-view image extraction method based on the three-color mask PIV technology provided by the present invention includes: acquiring the original mosaic image through a three-color mask single-color camera; constructing a neural network model composed of an image feature extraction module and an image reconstruction module; decomposing the original mosaic image into four channels; the four channels respectively represent different color information; performing Gaussian blur operations on the four channels respectively to obtain a preprocessed image; performing a first feature extraction on the preprocessed image through the initial convolution layer built into the image feature extraction module to obtain a preliminary image; performing multiple convolution and activation operations on the preliminary image through an activation function to obtain a feature image; and reconstructing the feature image through the image reconstruction module to obtain a three-view image. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0032] Figure 1 This is a flow chart of a particle three-view image extraction method based on three-color mask PIV technology according to an embodiment of the present invention;

[0033] Figure 2 A structural diagram of a neural network model according to an embodiment of the present invention;

[0034] Figure 3 This is the Q value calculation result diagram;

[0035] Figure 4 It is the artificial synthetic flow field map of the real image;

[0036] Figure 5 An artificial synthetic flow field diagram of an embodiment of the present invention;

[0037] Figure 6 Artificially synthesized flow field map for high-quality interpolation extraction algorithm;

[0038] Figure 7 This is the artificial flow field map of the gradient-based threshold-free extraction algorithm. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] like Figure 1 As shown, the embodiment of the present invention provides a particle three-view image extraction method based on three-color mask PIV technology, comprising the following steps:

[0042] Step 100: Acquire an original mosaic image through a three-color mask single color camera.

[0043] Specifically, the three-color mask PIV technology in the three-color mask single color camera is used to obtain a Bayer format particle image, and the particle image is input into the neural network model as the original mosaic image.

[0044] Step 200: Construct a neural network model consisting of an image feature extraction module and an image reconstruction module; the image feature extraction module is connected to the image reconstruction module; the neural network model is constructed according to the U-Net++ structure. The neural network model structure is as follows Figure 2 shown.

[0045] Specifically, the image feature extraction module includes multiple levels of Gaussian smoothing units ( Figure 2 The red arrow part) and multiple levels of feature extraction units ( Figure 2 The green box part); the Gaussian smoothing unit is connected to the feature extraction unit, and the feature extraction unit is connected to the image reconstruction module. The Gaussian smoothing unit is used to perform Gaussian blur processing on the original mosaic image; the feature extraction unit is used to extract the features of the original mosaic image.

[0046] More specifically, the Gaussian smoothing unit has four built-in color channels, namely, 1 R channel, 2 G channels and 1 B channel, and the four channels represent different color information of the original mosaic image. The Gaussian smoothing unit also has a built-in blur kernel for Gaussian blurring, and the size of the blur kernel is 5×5.

[0047] The feature extraction unit has built-in initial convolution layers of different levels, with sizes of 3×3, 5×5 or 7×7. Different initial convolution layer sizes are selected according to different levels. The feature extraction unit also has multiple operation layers built in. The unit structure of the feature extraction unit is a dense connection structure (Dense Unit), which enables multiple operation layers in the feature extraction unit to exist in the form of sequential connection between layers and cross-layer connection.

[0048] Specifically, the image reconstruction module includes: multiple levels of splicing units ( Figure 2 White circle part), multiple levels of reconstruction units ( Figure 2 The blue box part) and multiple levels of upsampling layers ( Figure 2 The blue arrow part); the splicing unit is connected to the feature extraction unit and the reconstruction unit respectively, and the reconstruction unit is connected to the upsampling layer; the splicing unit is used to splice the feature images; the reconstruction unit is composed of multiple convolutional layers and activation layers, and is used to perform convolution reconstruction operations on the feature images; the upsampling layer is used to restore the resolution of the feature image.

[0049] Specifically, constructing a neural network model composed of an image feature extraction module and an image reconstruction module also includes: training the neural network model through a final loss function; the specific steps are: using a structural similarity loss function to perform a first iteration on the neural network model to obtain a first iteration model; the number of iterations of the first iteration is 10 times; and then using a mean square error loss function to perform a second iteration on the first iteration model to obtain a trained neural network model.

[0050] Furthermore, the image reconstruction module has three output layers, whose outputs are represented by L1, L2 and L3 respectively. Among them, L1 is the low-level output of the model, corresponding to the image obtained after the least number of iterations. At this time, the model has not fully combined the features of images at different levels, so that the resolution of L1 is high, the image details are not rich enough, and the preliminary information of the image is retained. L2 is the intermediate output of the model. Compared with L1, it has undergone more iterations and contains richer image information and details, and has better visual effects than L1. L3 is the high-level output of the model and as the final output of the model, it has undergone more iterations than L2, and has fully extracted features and reconstructed the original mosaic image. The image quality is extremely high and contains the richest details.

[0051] Furthermore, the final loss function includes: structural similarity loss function and mean square error loss function; the expression of the final loss function is: Among them, L is the final loss function, is the structural similarity loss function value of the first output layer of the image reconstruction module, is the structural similarity loss function value of the second output layer of the image reconstruction module, is the structural similarity loss function value of the third output layer of the image reconstruction module, is the mean square error loss function value of the first output layer of the image reconstruction module, is the mean square error loss function value of the second output layer of the image reconstruction module, is the mean square error loss function value of the third output layer of the image reconstruction module, epoch is the number of iterations of the neural network model, SSIM is the structural similarity loss function, and MSE is the mean square error loss function.

[0052] Furthermore, the Structural Similarity Index (SSIM) analyzes the similarity of two images from three perspectives: brightness, contrast, and image structure, which can better reflect the human eye's perception of image quality. The SSIM index ranges from 0 to 1, where 1 means that the two images are exactly the same and 0 means that the two images are completely different. The calculation formula of the structural similarity loss function is:

[0053]

[0054] Among them, x and y are the corresponding blocks of the image, μ x and μ y are the average values ​​of x and y, respectively, x , σ y are the variances of x and y, σ xy is the covariance between x and y, C1 and C2 are constants, and SSIM is the structural similarity loss function.

[0055] Furthermore, the mean square error (MSE) of the image is used to measure the average pixel difference between two images. Specifically, it is obtained by calculating the square of the difference of each pixel in the image and then averaging it. The smaller the MSE, the more similar the two images are; the larger the MSE, the greater the difference between the two images. The calculation formula of the mean square error loss function is:

[0056]

[0057] Wherein, I and K are images to be compared, m and n are the width and height of the image respectively, and I(i, j) and K(i, j) are the pixel values ​​at position (i, j).

[0058] Step 300: Perform data processing operations on the original mosaic image through the image feature extraction module to obtain a feature image. The specific steps include:

[0059] Step 301: Decompose the original mosaic image into four channels. Further, the original mosaic image is decomposed into four channels (RGGB) by a Gaussian smoothing unit, so that each channel represents a color information in the original mosaic image.

[0060] Specifically, by decomposing the color information in the original mosaic image, the extracted color information features are made more specific and distinguishable, thereby improving the accuracy of the model.

[0061] Step 302: Gaussian blur operation is performed on the four channels respectively to obtain a preprocessed image. Further, Gaussian blur operation is performed on the four channels respectively using blur kernel, and Gaussian blur is performed for different times according to different levels of the image in the channel to obtain a preprocessed image.

[0062] Specifically, Gaussian blurring replaces the traditional downsampling operation, which has a similar effect to that of a Gaussian pyramid, but the image size remains unchanged after Gaussian blurring. Gaussian blurring expands the receptive field of each level, thereby fully extracting features without losing image information.

[0063] Step 303: Perform a first feature extraction on the preprocessed image through the initial convolution layer built into the image feature extraction module to obtain a preliminary image. Further, select a corresponding initial convolution layer size according to different levels of the preprocessed image, perform a first feature extraction operation on the preprocessed image through the initial convolution layer, and obtain a preliminary image.

[0064] Step 304: Perform multiple convolution and activation operations on the preliminary image through the activation function to obtain a feature image. Further, perform a depth-separable convolution operation on the preliminary image through multiple operation layers in the feature extraction unit to reduce the amount of calculation and the number of parameters of the network model. Then use the PReLU activation function to activate the image after the depth-separable convolution to complete the feature extraction operation to enhance the nonlinear expression ability of the network model. Finally, the activated image is used as the feature image.

[0065] Specifically, the output of each preceding operation layer in the feature extraction unit is used as the input of the next operation layer, so that each operation layer fully utilizes the features of the preceding operation layer when performing feature extraction, improves the efficiency of feature extraction, and reduces the parameters of the network model. The cross-layer connection of the operation layers in the feature extraction unit enables each operation layer to directly utilize the features of the previous layers, while reducing the number of network parameters, ensuring the depth and breadth of feature extraction and improving the integrity of feature extraction. Each operation layer has a built-in PReLU activation function.

[0066] Step 400: Reconstruct the feature image through an image reconstruction module to obtain a three-view image. The specific steps include:

[0067] First, the feature images extracted from different operation layers are spliced ​​through a splicing unit to obtain a spliced ​​image, so that the features of each operation layer can directly affect the result of image reconstruction, thereby improving the integrity and accuracy of the feature image.

[0068] Secondly, the stitched image is convolved and integrated layer by layer through the convolution layer and activation layer in the reconstruction unit and feature enhancement is performed to obtain an integrated image. By further convolution integration of the stitched image, the image reconstruction of the original mosaic image is completed, the close connection of each feature in the image is ensured, and the degree of retention of color information in the image is improved.

[0069] Finally, the integrated image is upsampled through the upsampling layer, and the resolution of the integrated image is adjusted to the resolution of the original mosaic image to obtain a high-resolution three-view image. The resolution adjustment makes the image details in the integrated image clearer.

[0070] The embodiment of the present invention uses a resolution of 1000×1000 pixels. 2 The mosaic image captured by a three-color mask single-color camera with a pixel size of 0.004 mm and a radius of 32 mm for the circumscribed circle of the three mask holes centers is used as an example for explanation.

[0071] First, a volume calibration image is generated using ray tracing technology, and the three-dimensional tracer particles are projected onto a two-dimensional image plane to obtain three viewing angle mapping matrices of the three-color mask. Subsequently, a group of particles are randomly generated in the measurement volume, and the projection of the particles at each viewing angle is calculated by the camera mapping matrix to determine the image coordinates of the center of the particle. Finally, a Gaussian distribution is applied to these coordinates to generate a particle image with a diameter of about 3 pixels. The above process is repeated at each viewing angle to obtain a three-color mask full-color particle image with a particle concentration of 0.05PPP (particles per-pixel). When training the network model, 5,000 simulated particle images were selected as network inputs, and 300 iterations were performed to finally obtain a demosaiced network model. The Bayer format mosaic image is demosaiced using the U-Net++-based extraction algorithm, the high-quality interpolation extraction algorithm (HQLI), and the gradient-based threshold-free extraction algorithm (GBTF) in the method of the present invention, and CPSNR and SSIM are used to represent the color reconstruction quality of the three algorithms. The results are shown in Table 1:

[0072] Table 1 De-mosaicing quality comparison table

[0073]

[0074] The residual analysis of the demosaiced particle image includes the calculation and analysis of the mean and standard deviation of the image. The results are shown in Table 2:

[0075] Table 2 Comparison of residual analysis results

[0076]

[0077] The results show that the residual mean and standard deviation of the U-Net++ algorithm in the demosaiced particle image are lower than those of the other two algorithms, indicating that the U-Net++ demosaicing algorithm is superior to the other two algorithms in terms of stability.

[0078] This example also introduces random noise of different intensities, 10%, 25%, and 40%, respectively. These noise intensities are expressed as percentages of the peak intensity of the particles. First, the grayscale values ​​of the images in the noisy dataset are averaged and used as the background for denoising. Subsequently, the denoised images are demosaiced, and the results are shown in Table 3:

[0079] Table 3 Comparison of the impact of noise on algorithm performance

[0080]

[0081] The results show that as the noise intensity increases, the SSIM value of the U-Net++-based extraction algorithm decreases, but it is still better than the high-quality interpolation extraction algorithm and the gradient-based threshold-free extraction algorithm.

[0082] This example uses the MLOS-SMART (Multiplicative Light of Sight-Simultaneous MART) algorithm to perform 3D reconstruction on the extracted particle three-view images. The 3D reconstruction quality Q value of the particle three-view images extracted by the above three algorithms is obtained. The Q value indicates the similarity between the particle reconstruction distribution matrix and the real distribution matrix. The calculation formula is:

[0083]

[0084] Among them, E0 is the distribution matrix of the artificially generated particle field, and E1 is the distribution matrix of the particles reconstructed by MLOS-SMART. The Q value calculation results of the three algorithms are as follows Figure 3 As shown in the figure, GroundTruth is the reconstruction quality Q value obtained after MLOS-SMART three-dimensional reconstruction. It can be seen that the extraction algorithm based on U-Net++ is closer to the reconstruction quality of full-color three-view images than the traditional high-quality interpolation extraction algorithm and the gradient-based threshold-free extraction algorithm.

[0085] This example also uses artificial Gaussian vortex ring velocity field simulation experiments to analyze the velocity field measurement accuracy. The artificial synthetic flow field of the real image and the artificial synthetic flow fields of the above three algorithms are shown in Figure 2. Figures 4 to 7 As shown, it can be seen that the artificial synthetic flow field of the method of the present invention is the closest to the artificial synthetic flow field of the real image and is the most stable, which greatly reduces the influence of the extraction quality of the three-view image on the measurement error.

[0086] The beneficial effects of the present invention are as follows:

[0087] 1) Improved the efficiency of feature extraction process;

[0088] 2) Improved the integrity and accuracy of three-view images;

[0089] 3) The quality of the extracted three-view images is improved;

[0090] 4) The influence of the extraction quality of three-view images on the measurement error is reduced.

[0091] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0092] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A particle three-view image extraction method based on three-color mask PIV technology, characterized in that: The steps include: The original mosaic image is obtained by a three-color mask single color camera; Constructing a neural network model consisting of an image feature extraction module and an image reconstruction module; the image feature extraction module is connected to the image reconstruction module; the neural network model is constructed according to the U-Net++ structure; The image feature extraction module is used to perform data processing on the original mosaic image to obtain a feature image. The specific steps of performing data processing on the original mosaic image by the image feature extraction module include: Decomposing the original mosaic image into four channels; the four channels respectively represent different color information; Performing Gaussian blur operation on the four channels respectively to obtain a preprocessed image; Performing a first feature extraction on the preprocessed image through the initial convolution layer built into the image feature extraction module to obtain a preliminary image; Performing multiple convolution and activation operations on the preliminary image through an activation function to obtain the feature image; The image reconstruction module is used to reconstruct the characteristic image to obtain a three-viewing angle image.

2. The particle three-view image extraction method based on three-color mask PIV technology according to claim 1, characterized in that: The image feature extraction module includes a plurality of Gaussian smoothing units and a plurality of feature extraction units; the Gaussian smoothing units are connected to the feature extraction units, and the feature extraction units are connected to the image reconstruction module; The Gaussian smoothing unit is used to perform Gaussian blur processing on the original mosaic image; and the feature extraction unit is used to extract features of the original mosaic image.

3. The particle three-view image extraction method based on three-color mask PIV technology according to claim 2 is characterized in that: The image reconstruction module includes: a plurality of splicing units, a plurality of reconstruction units and a plurality of upsampling layers; the splicing units are respectively connected to the feature extraction unit and the reconstruction unit, and the reconstruction unit is connected to the upsampling layer; The splicing unit is used to splice the feature images; the reconstruction unit is used to perform a convolution reconstruction operation on the feature images; and the upsampling layer is used to restore the resolution of the feature images.

4. The particle three-view image extraction method based on three-color mask PIV technology according to claim 1, characterized in that: The image reconstruction module is used to reconstruct the feature image to obtain a three-view image, and the specific steps include: Performing a stitching operation on the feature images to obtain a stitched image; Integrating and enhancing the features of the stitched images to obtain an integrated image; An upsampling operation is performed on the integrated image to obtain the three-view image.

5. The particle three-view image extraction method based on three-color mask PIV technology according to claim 1, characterized in that: Construct a neural network model consisting of an image feature extraction module and an image reconstruction module, and also include: The neural network model is trained by a final loss function; the final loss function includes: a structural similarity loss function and a mean square error loss function; the expression of the final loss function is: Among them, L is the final loss function, is the structural similarity loss function value of the first output layer of the image reconstruction module, is the structural similarity loss function value of the second output layer of the image reconstruction module, is the structural similarity loss function value of the third output layer of the image reconstruction module, is the mean square error loss function value of the first output layer of the image reconstruction module, is the mean square error loss function value of the second output layer of the image reconstruction module, is the mean square error loss function value of the third output layer of the image reconstruction module, epoch is the number of iterations of the neural network model, SSIM is the structural similarity loss function, and MSE is the mean square error loss function.

6. The particle three-view image extraction method based on three-color mask PIV technology according to claim 5, characterized in that: The neural network model is trained by the final loss function, and the specific steps include: Performing a first iteration on the neural network model using the structural similarity loss function to obtain a first iteration model; the number of iterations of the first iteration is 10 times; The first iterative model is iterated for a second time using the mean square error loss function to obtain a trained neural network model.

7. The particle three-view image extraction method based on three-color mask PIV technology according to claim 5, characterized in that: The calculation formula of the structural similarity loss function is: Among them, x and y are the corresponding blocks of the image, μ x and μ y are the average values ​​of x and y, respectively, x , σ y are the variances of x and y, σ xy is the covariance between x and y, C1 and C2 are constants, and SSIM is the structural similarity loss function.

8. The particle three-view image extraction method based on three-color mask PIV technology according to claim 5, characterized in that: The calculation formula of the mean square error loss function is: Wherein, I and K are images to be compared, m and n are the width and height of the image respectively, and I(i, j) and K(i, j) are the pixel values ​​at position (i, j).

Images