Three-color mask single-color camera particle image velocity measurement particle three-dimensional distribution reconstruction method

Abstract

The invention belongs to the technical field of three-dimensional three-component velocity measurement, and provides a three-color mask single-color camera particle image velocity measurement particle three-dimensional distribution reconstruction method, which comprises the steps of color particle image shooting, color channel separation, preliminary three-dimensional reconstruction and three-dimensional reconstruction refinement by using a convolutional neural network model. According to the method, three-dimensional reconstruction is performed through the convolutional neural network model, so that the spatial position and the intensity distribution of the particles are more in line with the real situation, the particle elongation effect is greatly reduced, the calculation efficiency is improved, and the method has higher resolution in the depth-of-field direction.

Description

Technical Field

[0001] The invention relates to the technical field of three-dimensional three-component velocity measurement, and in particular to a method for reconstructing three-dimensional distribution of particles by particle image velocity measurement using a three-color mask single-color camera. Background Art

[0002] Particle image velocimetry is a non-contact flow field measurement technology that can accurately obtain transient quantitative information of the entire flow field and is widely used in the display and measurement of flow field structures. Tomographic particle image velocimetry is a three-dimensional three-component velocity measurement technology that can measure transient three-dimensional unsteady flow field velocities. This technology uses 3 to 6 cameras to record tracer particles in the measured area of ​​the fluid from different angles, and calculates the velocity field information through reconstruction algorithms and three-dimensional cross-correlation algorithms. Among them, three-dimensional reconstruction of particles is an important step. This step calculates the intensity distribution of particles in three-dimensional space, and the quality of the results affects the measurement accuracy of the final flow field. Because the number of cameras required for tomographic particle image velocimetry is large, it is inconvenient to arrange them when the number and size of optical measurement windows are limited. Three-color mask particle image velocimetry is a single-camera testing technology. A three-color mask device is installed in front of the lens of a single color camera. It consists of three narrow-band filters with different central wavelengths and a black mask. The black mask has light holes arranged in a certain pattern to produce a fixed parallax. The imaging optical path is changed by this device, and the parallax information is encoded as color information. In a single exposure, particle image information from three perspectives can be captured simultaneously and recorded in different color channels of the color particle image. After the color channel separation operation is performed on the three-color mask color particle image and the three-perspective images are extracted, the particle spatial distribution needs to be reconstructed using a method similar to tomographic particle image velocimetry, and finally the three-dimensional velocity field is obtained according to the cross-correlation algorithm.

[0003] However, due to the small parallax of single-camera imaging, reconstructing the three-color mask particle image using traditional reconstruction methods such as the multiplicative line-of-sight synchronous algebraic reconstruction technology suitable for tomographic particle image velocimetry requires hundreds of large iterative calculations. In addition, particles that should theoretically be reconstructed as spherical will actually produce a significant elongation effect along the depth of field direction. The particle shape appears to be elongated and appears to be spindle-shaped or even needle-shaped, which seriously affects the reconstruction accuracy and reduces the spatial measurement resolution in the depth of field direction. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a method for reconstructing the three-dimensional distribution of particles by particle image velocimetry using a three-color mask single-color camera. The three-dimensional reconstruction is performed through a convolutional neural network model so that the spatial position and intensity distribution of the particles are more consistent with the actual situation, thereby reducing the particle elongation effect, improving the calculation efficiency and the resolution in the depth of field direction.

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

[0006] A method for reconstructing three-dimensional distribution of particles by particle image velocimetry using a three-color mask single-color camera, comprising:

[0007] Take color particle images;

[0008] Using a demosaicing algorithm to separate the color channels of the color particle image to obtain a three-view particle grayscale image;

[0009] Using the MLOS algorithm to perform preliminary three-dimensional reconstruction on the three-view particle grayscale image to obtain an initial voxel field;

[0010] Using the trained convolutional neural network model, the particles in the initial voxel field are refined and three-dimensionally reconstructed into a circle to obtain a refined voxel field;

[0011] The training process of the convolutional neural network model includes:

[0012] Setting optical parameters, and calculating a projection transformation matrix according to the optical parameters;

[0013] Setting several particle concentrations, and randomly generating equal-amount particle spatial intensity voxel fields according to the particle concentrations;

[0014] The projection transformation matrix is ​​used to convert the equal-amount particle spatial intensity voxel field into a grayscale two-dimensional image of the particles to be trained;

[0015] Using the MLOS algorithm to perform preliminary three-dimensional reconstruction on the grayscale two-dimensional image of the particles to be trained, to obtain a voxel field to be trained;

[0016] Randomly dividing the voxel field to be trained into a training set, a validation set and a test set according to a preset ratio;

[0017] Constructing a network model; the network model includes several hidden network layers; the hidden network layers include: a 3D convolution layer, a batch normalization layer and an activation function layer;

[0018] Construct loss function;

[0019] The network model is trained using the Adam optimizer and the exponential moving average method according to the training set and the verification set to obtain a trained model to be tested;

[0020] Using the test set to test the model to be tested to obtain a test result;

[0021] When the test result meets the preset standard, the model to be tested is output to obtain the convolutional neural network model.

[0022] Preferably, the optical parameters include: pixel size, focal length, nominal aperture, particle imaging area, measurement area size, single mask hole diameter, magnification, and radius of the circumscribed circle of the mask hole center.

[0023] Preferably, the loss function is:

[0024]

[0025] Wherein, Loss is the loss function; E 0i and E predi are respectively the intensity values ​​of the equal-particle spatial intensity voxel field and the predicted voxel field obtained by the network model at position i.

[0026] Preferably, the calculation formula of the test result is:

[0027]

[0028] Wherein, Q is the test result; E 1i is the intensity value of the reconstructed voxel obtained by the model to be tested at position i.

[0029] The present invention discloses the following technical effects:

[0030] The present invention provides a method for reconstructing the three-dimensional distribution of particles by measuring particle images using a three-color mask single-color camera. The three-dimensional reconstruction is performed through a convolutional neural network model, which solves the defects of the prior art of strong particle elongation effect and poor computational efficiency, and realizes fast and accurate reconstruction of the three-dimensional distribution of particles. 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 A schematic diagram of a process for reconstructing a three-dimensional distribution of particles using a three-color mask single-color camera particle image velocimetry provided in an embodiment of the present invention;

[0033] Figure 2 A structural diagram of a convolutional neural network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] 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.

[0035] The purpose of the present invention is to provide a method for reconstructing the three-dimensional distribution of particles by particle image velocimetry using a three-color mask single-color camera. The three-dimensional reconstruction is performed through a convolutional neural network model so that the spatial position and intensity distribution of the particles are more consistent with the actual situation, thereby reducing the particle elongation effect, improving the calculation efficiency and the resolution in the depth of field direction.

[0036] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] Figure 1 The schematic diagram of the process of reconstructing the three-dimensional distribution of particles by using a three-color mask single color camera particle image velocimetry according to an embodiment of the present invention is as follows: Figure 1 As shown, the present invention provides a method for reconstructing three-dimensional distribution of particles by particle image velocimetry using a three-color mask single-color camera, comprising:

[0038] Step 100: photographing a color particle image;

[0039] Step 200: using a demosaicing algorithm to separate the color channels of the color particle image to obtain a three-view particle grayscale image;

[0040] Step 300: using the MLOS algorithm to perform preliminary three-dimensional reconstruction on the three-view particle grayscale image to obtain an initial voxel field;

[0041] Step 400: using the trained convolutional neural network model to refine the particles in the initial voxel field into a three-dimensional circle to obtain a refined voxel field;

[0042] The training process of the convolutional neural network model includes:

[0043] Set optical parameters and calculate the projection transformation matrix based on the optical parameters;

[0044] Set several particle concentrations and randomly generate equal-amount particle spatial intensity voxel fields according to the particle concentrations;

[0045] The equal-amount particle spatial intensity voxel field is converted into a grayscale two-dimensional image of the particles to be trained using a projection transformation matrix;

[0046] The MLOS algorithm is used to perform preliminary 3D reconstruction on the grayscale 2D image of the particles to be trained, and the voxel field to be trained is obtained;

[0047] The voxel field to be trained is randomly divided into a training set, a validation set and a test set according to a preset ratio;

[0048] Construct a network model; the network model includes several hidden network layers; the hidden network layers include: 3D convolution layer, batch normalization layer and activation function layer;

[0049] Construct loss function;

[0050] The network model is trained using the Adam optimizer and the exponential moving average method according to the training set and the validation set to obtain a trained model to be tested;

[0051] Use the test set to test the model to be tested and obtain the test results;

[0052] When the test results meet the preset standards, the model to be tested is output to obtain a convolutional neural network model.

[0053] Optionally, the optical parameters include: pixel size, focal length, nominal aperture, particle imaging area, measurement area size, single mask hole diameter, magnification, and radius of the circumscribed circle of the mask hole center.

[0054] Specifically, the loss function is:

[0055]

[0056] Among them, Loss is the loss function; and They are the intensity values ​​of the equal-particle spatial intensity voxel field and the predicted voxel field obtained by the network model at position i.

[0057] Furthermore, the calculation formula of the test result is:

[0058]

[0059] Among them, Q is the test result; is the intensity value of the reconstructed voxel obtained by the model to be tested at position i.

[0060] Specifically, the particles reconstructed using a neural network can greatly reduce the particle elongation effect, reconstruct the particle shape into a circle, and the particle spatial position and intensity distribution are basically consistent with the actual situation, thus having a higher resolution in the depth of field direction. The test set data is used to verify the reconstruction prediction performance of the neural network model, and the reconstruction quality factor (test results) is used as the evaluation index to evaluate the reconstruction quality under different parallax conditions (here the particle concentration is 0.05ppp), and compared with the MLOS-SMART method (the calculation results of the reconstruction quality factor are shown in Table 1).

[0061]

[0062] The value of Q ranges from 0 to 1. The higher the Q value, the better the reconstruction quality. Quality evaluation of reconstruction results using different algorithms under different parallaxes and fixed particle concentration.

[0063] Table 1

[0064]

[0065]

[0066] Furthermore, the parallax is fixed at a magnification of -1, the mask hole spacing is 32 mm, and test samples with different particle concentrations are selected. The reconstruction quality factor is used as the evaluation index to evaluate the reconstruction quality under different particle concentrations, and the MLOS-SMART method is used for comparison. The quality evaluation of the reconstruction results using different algorithms under fixed parallax and different particle concentrations is shown in Table 2.

[0067] Table 2

[0068]

[0069] Preferably, the use of neural networks can greatly improve the particle reconstruction speed, and the effect is stable and has higher reconstruction efficiency; in contrast, the reconstruction speed of the MLOS-SMART method will be affected by factors such as the number of iterations, particle concentration and parallax size.

[0070] Specifically, the specific process of the particle three-dimensional distribution reconstruction method of particle image velocimetry using a three-color mask single-color camera is as follows:

[0071] S1: Set optical parameters; according to the camera parameter arrangement conditions used in the experiment, set the optical parameters such as pixel size, focal length, nominal aperture, particle imaging area, measurement area size, single mask hole diameter, magnification and mask hole center circumscribed circle radius (mask hole spacing). Among them, the combination of magnification and mask hole spacing determines the shooting parallax. The magnification is defined as follows:

[0072] M=-S I / S O

[0073] Where M represents the magnification, S I With S O They represent image distance and object distance respectively.

[0074] S2: Calculate the projection transformation matrix at each imaging viewing angle according to the optical parameters.

[0075] S3: Set multiple particle concentrations, randomly generate particle space coordinates according to the particle concentrations, and assign values ​​to the spatial voxel field with Gaussian intensity distribution. The obtained voxel field is the real voxel field E0, which is used as the label data sample of the neural network. The number of samples under each particle concentration should be consistent;

[0076] S4: For each voxel field, a two-dimensional grayscale image of the particles at each viewing angle is generated by combining the projection transformation matrix, simulating the full-frame particle images of three viewing angles obtained after separating the color channels;

[0077] S5: Use the MLOS algorithm to preliminarily reconstruct the particle image and obtain the voxel field E with the initial value. MLOS , as the input data sample of the neural network; reference Figure 2 , the voxel field size of the training sample is 200×128×128 voxels, and the number of voxels in the depth direction is 200;

[0078] S6: After grouping and shuffling the data samples of different concentrations, divide them into training set, validation set and test set in proportion.

[0079] S7: Network training. Using convolutional neural network (ref. Figure 2 ), each hidden layer consists of a 3D convolution layer, a batch normalization layer, and an activation function layer; the depth of the 3D convolution layer is set larger than the height and width, so that it can have a larger receptive field in the depth direction and learn more depth-direction features; except for the last hidden layer, the activation function uses the ReLU function, and the activation function of the last hidden layer uses the Sigmoid function. Figure 2 , 10 hidden layers with 16 channels are set in the neural network, and the number of channels in the remaining hidden layers is 1; the size of the voxel field in the network remains unchanged in each layer. The neural network uses Dice Loss as the loss function, and its formula is as follows:

[0080]

[0081] According to the loss function, the Adam optimization method is used, and the exponential moving average training strategy is adopted to optimize and update the network parameters.

[0082] S8: Take three-color mask PIV experimental particle images, use the demosaicing algorithm to separate the color channels, and obtain particle images from three perspectives.

[0083] S9: Reconstruct the particle image using the MLOS algorithm.

[0084] S10: Use the trained neural network model to refine and reconstruct the voxel field obtained by the MLOS algorithm. Obtain a refined voxel field, improve the reconstructed shape and intensity distribution of three-dimensional particles, reduce the particle elongation effect, and thus improve the reconstruction accuracy. At the same time, this step consumes less computing resources and has a fast calculation speed.

[0085] The beneficial effects of the present invention are as follows:

[0086] The present invention uses a convolutional neural network model to perform three-dimensional reconstruction, making the particle spatial position and intensity distribution more consistent with the actual situation, greatly reducing the particle elongation effect, improving the calculation efficiency, and having a higher resolution in the depth of field direction.

[0087] 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.

[0088] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea 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 method for reconstructing three-dimensional distribution of particles by particle image velocimetry using a three-color mask single-color camera, characterized in that: include: Take color particle images; Using a demosaicing algorithm to separate the color channels of the color particle image to obtain a three-view particle grayscale image; Using the MLOS algorithm to perform preliminary three-dimensional reconstruction on the three-view particle grayscale image to obtain an initial voxel field; Using the trained convolutional neural network model, the particles in the initial voxel field are refined and three-dimensionally reconstructed into a circle to obtain a refined voxel field; The training process of the convolutional neural network model includes: Setting optical parameters, and calculating a projection transformation matrix according to the optical parameters; Setting several particle concentrations, and randomly generating equal-amount particle spatial intensity voxel fields according to the particle concentrations; The projection transformation matrix is ​​used to convert the equal-amount particle spatial intensity voxel field into a grayscale two-dimensional image of the particles to be trained; Using the MLOS algorithm to perform preliminary three-dimensional reconstruction on the grayscale two-dimensional image of the particles to be trained, to obtain a voxel field to be trained; Randomly dividing the voxel field to be trained into a training set, a validation set and a test set according to a preset ratio; Constructing a network model; the network model includes several hidden network layers; the hidden network layers include: a 3D convolution layer, a batch normalization layer and an activation function layer; Construct loss function; The network model is trained using the Adam optimizer and the exponential moving average method according to the training set and the verification set to obtain a trained model to be tested; Using the test set to test the model to be tested to obtain a test result; When the test result meets the preset standard, the model to be tested is output to obtain the convolutional neural network model.

2. The method for reconstructing three-dimensional distribution of particles by particle image velocimetry using a three-color mask single-color camera according to claim 1, characterized in that: The optical parameters include: pixel size, focal length, nominal aperture, particle imaging area, measurement area size, single mask hole diameter, magnification, and radius of the circumscribed circle of the mask hole center.

3. The method for reconstructing three-dimensional distribution of particles by particle image velocimetry using a three-color mask single-color camera according to claim 1, characterized in that: The loss function is: Wherein, Loss is the loss function; and are respectively the intensity values ​​of the equal-particle spatial intensity voxel field and the predicted voxel field obtained by the network model at position i.

4. The method for reconstructing three-dimensional distribution of particles by particle image velocimetry using a three-color mask single-color camera according to claim 3, characterized in that: The calculation formula of the test result is: Wherein, Q is the test result; is the intensity value of the reconstructed voxel obtained by the model to be tested at position i.

Images