Capacitance tomography image reconstruction method, model, device and storage medium

Abstract

The present invention provides a capacitance tomography image reconstruction method, model, device and storage medium, wherein the capacitance tomography image super-resolution reconstruction method comprises: using a first-level neural network to perform nonlinear image reconstruction from a capacitance vector to a low-resolution medium distribution, wherein the input of the first-level neural network is a capacitance matrix obtained by rearranging the normalized capacitance vector; the output of the first-level neural network is S 1×1 dimensional medium distribution column vector, S 1 is the number of pixel units of the low-resolution medium distribution image; the second-level neural network is used to perform super-resolution reconstruction of the image from low-resolution medium distribution to high-resolution medium distribution; the input of the second-level neural network is the output of the first-level neural network, and the output is S 2×1 dimensional medium distribution column vector, S 2 is the number of pixel units of the high-resolution medium distribution image. The present invention can achieve high spatial resolution ECT image reconstruction while greatly reducing the time required to establish a large sample capacity data set.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical capacitance tomography, and in particular to an electrical capacitance tomography image reconstruction method, model, device and storage medium. Background Art

[0002] For electrical capacitance tomography (ECT) systems, spatial resolution refers to the size of the smallest distinguishable pixel unit in the reconstructed image. The smaller the pixel unit and the greater its number, the higher the spatial resolution of the system.

[0003] Capacitance tomography technology uses the different dielectric constants of multiphase media to reflect the distribution changes of multiphase media in the pipeline through the change of array electrode capacitance, thereby realizing the reconstruction of multiphase flow visualization images. ECT technology has the advantages of non-invasiveness, simple sensor structure, low cost, fast response speed, wide application range and high safety performance, and is widely used in the detection of multiphase flow processes of insulating media.

[0004] However, compared with other process tomography techniques, such as optical tomography and radiation tomography, the spatial resolution of capacitance tomography is low and the imaging quality is poor. There are two main factors that lead to the low spatial resolution of ECT: first, the number of independent capacitance measurement values ​​is far less than the number of pixel units in the reconstructed image, that is, the "morbid" characteristic; second, the linear simplification of the nonlinear relationship between the capacitance vector and the medium distribution by the traditional reconstruction algorithm causes the loss of projection information. In order to improve the spatial resolution of the ECT system, the number of independent capacitance value measurements can be increased by increasing the number of capacitance sensor plates. However, the increase in the number of electrodes will lead to a smaller distance between adjacent electrodes, an enhanced electrode edge effect, and more electric field lines crossing the field boundary, resulting in an increase in the sensitivity of the field boundary and a decrease in the sensitivity of the field center, which is not conducive to improving the imaging quality. Therefore, the current ECT system generally uses sensors with 8-electrode, 12-electrode or 16-electrode structures, and it is difficult to improve the spatial resolution from the perspective of increasing the number of capacitance measurement values.

[0005] At present, the algorithms used for ECT image reconstruction can be mainly divided into two categories: linear models and nonlinear models. The linear model algorithm linearly simplifies the nonlinear mapping between the capacitance and the dielectric distribution, and describes the physical relationship between the two through the sensitive field matrix. The linear model algorithm can be divided into three categories: direct algorithm, iterative algorithm and intelligent optimization algorithm. The direct algorithm directly uses the pseudo-inverse of the sensitive field matrix for back projection, such as the LBP algorithm and the Tikhonov algorithm. This type of algorithm has a small amount of calculation and fast reconstruction speed, but low reconstruction accuracy. The iterative algorithm is an approximate solution of the sensitive field inverse matrix through linear iteration, such as the Landweber algorithm and the conjugate gradient method. The accuracy of this type of algorithm is improved, but the reconstruction speed is slow. The intelligent optimization algorithm converts the inverse problem into an optimization problem, and uses the optimization algorithm to globally search for the optimal approximation of the sensitive field inverse matrix, such as the simulated annealing algorithm and the genetic algorithm. This type of algorithm has high accuracy, but the results are difficult to converge, and have great limitations in practical applications. The nonlinear model algorithm realizes image reconstruction by directly fitting the nonlinear relationship between capacitance and medium distribution. At present, the main nonlinear model algorithms are deep learning algorithms, such as deep fully connected networks, autoencoder networks, convolutional neural networks, etc. This type of algorithm has high reconstruction accuracy and fast reconstruction speed, and avoids the information loss caused by linear simplification, which is conducive to improving the reconstruction resolution.

[0006] The deep learning model learns the characteristics of the data set to predict the samples that have not been learned, so the capacity and richness of the data set have a crucial impact on the performance of the deep learning model. At present, the data sets used to train deep learning models are generally generated by numerical simulation. First, a certain number of medium distribution images are generated according to the preset spatial resolution; then, the capacitance vector corresponding to each medium distribution is calculated through the ECT finite element forward model to form a complete "capacitance vector-medium distribution" data set. The training of deep learning models depends on data sets with large sample capacity, but as the resolution of the reconstructed image increases, the time required for forward calculation of the capacitance vector increases greatly. Directly establishing a "capacitance vector-high-resolution medium distribution" data set will consume a lot of computing resources and time. Summary of the invention

[0007] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and to propose an electrical capacitance tomography image reconstruction method, model, device and storage medium that reduce the calculation time and difficulty.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0009] The present invention provides a method for super-resolution reconstruction of electrical capacitance tomography images, comprising:

[0010] The first-level neural network is used to perform nonlinear image reconstruction from capacitance vector to low-resolution medium distribution, where the input layer of the first-level neural network is the normalized capacitance vector rearranged. N × N dimensional capacitance matrix, N is the number of electrodes; the output of the first-level neural network is S 1×1 dimensional medium distribution column vector, S 1 is the number of pixel units of the low-resolution medium distribution image;

[0011] The second-level neural network is used to perform super-resolution reconstruction of images from low-resolution medium distribution to high-resolution medium distribution; the input of the second-level neural network is the output of the first-level neural network, and the output of the second-level neural network is S 2×1 dimensional medium distribution column vector, S 2 is the number of pixel units of the high-resolution medium distribution image.

[0012] The first-level neural network is a first-level convolutional neural network, including an input layer, an upsampling layer, a convolution layer, a fully connected layer, and an output layer; the input layer is obtained by rearranging the normalized capacitance vector N × N dimensional capacitance matrix; after the input layer, the input capacitance matrix is ​​unpooled and upsampled to expand the data dimension; the convolution layer uses two-dimensional convolution, setting m 1 n 1× n 1D convolution kernel, the convolution kernel step size is 1; the fully connected layer is located after the convolution layer, through l 1 neuron integrates the highly abstract features after convolution and finally outputs S 1×1 dimensional medium distribution column vector.

[0013] A Dropout layer is added after the convolutional layer and the fully connected layer of the first-level convolutional neural network to randomly delete some neurons and their connections during training to enhance the generalization ability of the network.

[0014] The second-level neural network is a second-level convolutional neural network, including an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer; the input layer is the reconstructed S 1×1 dimensional low-resolution medium distribution column vector; the convolution layer uses one-dimensional convolution, setting m 2 n 2×1 convolution kernel, convolution kernel step size is 1; pooling layer uses one-dimensional maximum pooling to shorten the length of column vector to 1 / 2 of the original; finally, through the fully connected layer l 2 neurons are connected to each node of the output layer, and the output S 2×1 dimensional medium distribution column vector.

[0015] A Dropout layer is set after the convolutional layer and the fully connected layer of the second-level convolutional neural network.

[0016] The present invention also provides an electrical capacitance tomography image reconstruction model, comprising:

[0017] The first-level neural network performs nonlinear image reconstruction from capacitance vector to low-resolution medium distribution, where the input layer of the first-level neural network is the normalized capacitance vector rearranged N × N dimensional capacitance matrix, N is the number of electrodes; the output of the first-level neural network is S 1×1 dimensional medium distribution column vector, S 1 is the number of pixel units of the low-resolution medium distribution image;

[0018] The second-level neural network performs super-resolution reconstruction of images from low-resolution medium distribution to high-resolution medium distribution; the input of the second-level neural network is the output of the first-level neural network, and the output of the second-level neural network is S 2×1 dimensional medium distribution column vector, S 2 is the number of pixel units of the high-resolution medium distribution image.

[0019] The present invention also provides a device, characterized in that it comprises:

[0020] one or more processors;

[0021] A memory for storing one or more programs;

[0022] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned capacitance tomography image reconstruction method.

[0023] The present invention also provides a storage medium having a computer program stored thereon, characterized in that the program implements the above-mentioned capacitance tomography image reconstruction method when executed by a processor.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] Constructing the "capacitance vector-medium distribution" data set by numerical simulation requires three steps: generating random medium distribution, matching grids with different resolutions, and calculating capacitance vector finite elements. Among them, the time required for the capacitance vector calculation step will greatly increase with the increase of spatial resolution; taking the generation of 10,000 samples as the benchmark, after the resolution is increased from 510 to 8053, the capacitance calculation time increases from 1682s to 43703s, an increase of about 2500%. Although high-performance computers can be used to construct high-resolution image data sets, the time-consuming problem of data set construction cannot be fundamentally solved, and it is necessary to find a solution to this problem from the perspective of the model. On the other hand, since the solution speed of the calculation model obtained after learning and training is not affected by the training process, the problem of low data construction efficiency has not received attention and attention. But in fact, reducing the construction time of the data set and ensuring the efficiency of the model are of great practical significance for the flexibility of measurement in engineering applications.

[0026] The capacitance tomography image super-resolution reconstruction method of the present invention is based on the capacitance tomography image super-resolution reconstruction model of the primary network and the secondary network, first completing the conversion of capacitance vector to low-resolution medium distribution image, and then completing the conversion of low-resolution medium distribution image to high-resolution medium distribution image, thereby greatly improving the spatial resolution of the ECT reconstructed image. At the same time, the method of the present invention only requires the finite element calculation of capacitance for the low-resolution image, avoiding the direct calculation of the capacitance vector corresponding to the high-resolution image, thereby greatly reducing the time required to establish a large sample capacity data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flow chart of the electrical capacitance tomography image reconstruction method of the present invention;

[0028] Figure 2 This is the network model structure for super-resolution reconstruction of electrical capacitance tomography images of the present invention. DETAILED DESCRIPTION

[0029] The capacitance tomography image reconstruction method proposed in the present invention has the following process: Figure 1 As shown: first, the "capacitance vector" is input into the first-level network, and after a series of nonlinear transformations, the "low-resolution medium distribution" is output, and the corresponding low-resolution image can be drawn; then the "low-resolution medium distribution" is input into the second-level network, and finally the "high-resolution medium distribution" and the corresponding high-resolution image can be output, thereby realizing image super-resolution reconstruction.

[0030] The model structure for realizing image super-resolution reconstruction is as follows: Figure 2 As shown, Figure 2 The middle and upper layers are the first-level networks. Figure 2 The middle and lower layers are the second-level neural network.

[0031] In one embodiment, both the first-level network and the second-level network can be convolutional neural networks.

[0032] The first-level convolutional neural network is mainly composed of input layer, upsampling layer, convolution layer, fully connected layer and output layer. The input layer is the normalized capacitance vector rearranged N × N dimensional capacitance matrix, N is the number of electrodes. After the input layer, the input capacitance matrix is ​​unpooled and upsampled to expand the data dimension and enhance the amount of information that the network can extract. The convolution layer uses two-dimensional convolution and sets m 1 n 1× n 1D convolution kernel, convolution kernel step size is 1. The fully connected layer is located after the convolution layer, through l 1 neuron integrates the highly abstract features after convolution and finally outputs S 1×1 dimensional medium distribution column vector, S 1 is the number of pixel units in the low-resolution medium distribution image. A Dropout layer is added after the convolutional layer and the fully connected layer to randomly delete some neurons and their connections during training to enhance the generalization ability of the network.

[0033] The second-level convolutional neural network is mainly composed of input layer, convolution layer, pooling layer, fully connected layer and output layer. The input layer is reconstructed by the first-level neural network. S 1×1 dimensional low-resolution medium distribution column vector. The convolution layer uses one-dimensional convolution and sets m 2 n 2×1 convolution kernel, convolution kernel step size is 1. The pooling layer uses one-dimensional maximum pooling to shorten the length of the column vector to 1 / 2 of the original. Finally, through the fully connected layer l 2 neurons are connected to each node of the output layer, and the output S 2×1 dimensional medium distribution column vector, S 2 is the number of pixel units of the high-resolution medium distribution image. A Dropout layer is also set after the convolutional layer and the fully connected layer.

[0034] For the two convolutional neural network models, it is necessary to establish a "capacitance vector-low-resolution medium distribution" data set and a "low-resolution medium distribution-high-resolution medium distribution" data set, and train the above two networks separately. After the training is completed, a complete high-resolution medium distribution prediction model can be formed.

[0035] Table 1 is a statistical diagram of the time required for each step to generate 10,000 samples at different spatial resolutions. It can be seen that as the spatial resolution increases, the time required for capacitance calculation increases greatly. If the construction of a data set of 100,000 samples is used as a benchmark, and the goal is to reconstruct an image with a spatial resolution of 8053 grid units, the data set construction time required by a single-level neural network is about 122 hours; and the method proposed in the present invention avoids direct capacitance calculation on high-resolution grids by adding a reconstruction network from low resolution to high resolution, and the required data set construction time can be reduced to about 6.5 hours at the lowest, with a maximum reduction of about 95%. Therefore, this method can effectively solve the problem of long construction time of high-resolution data sets.

[0036] Table 1 The time required for each step to generate 10,000 samples at different spatial resolutions

[0037] Number of grid cells 510 1004 2028 4032 8053 Flow pattern generation time (s) 122 122 122 122 122 Grid matching time (s) 189 182 189 206 225 Capacitance calculation time (s) 1682 2966 6811 16278 43703

[0038] The embodiment of the present invention also provides an electrical capacitance tomography image reconstruction model, such as Figure 2 As shown, including:

[0039] The first-level neural network performs nonlinear image reconstruction from capacitance vector to low-resolution medium distribution, where the input layer of the first-level neural network is the normalized capacitance vector rearranged N × N dimensional capacitance matrix, N is the number of electrodes; the output of the first-level neural network is S 1×1 dimensional medium distribution column vector, S 1 is the number of pixel units of the low-resolution medium distribution image;

[0040] The second-level neural network performs super-resolution reconstruction of images from low-resolution medium distribution to high-resolution medium distribution; the input of the second-level neural network is the output of the first-level neural network, and the output of the second-level neural network is S 2×1 dimensional medium distribution column vector, S 2 is the number of pixel units of the high-resolution medium distribution image.

[0041] The first-level neural network is a first-level convolutional neural network, including input layer, upsampling layer, convolution layer, fully connected layer and output layer; the input layer is the normalized capacitance vector rearranged N × N dimensional capacitance matrix; after the input layer, the input capacitance matrix is ​​unpooled and upsampled to expand the data dimension; the convolution layer uses two-dimensional convolution, setting m 1 n 1× n1D convolution kernel, the convolution kernel step size is 1; the fully connected layer is located after the convolution layer, through l 1 neuron integrates the highly abstract features after convolution and finally outputs S 1×1 dimensional medium distribution column vector.

[0042] The second-level neural network is a second-level convolutional neural network, including input layer, convolution layer, pooling layer, fully connected layer and output layer; the input layer is S 1×1 dimensional low-resolution medium distribution column vector; the convolution layer uses one-dimensional convolution, setting m 2 n 2×1 convolution kernel, convolution kernel step size is 1; pooling layer uses one-dimensional maximum pooling to shorten the length of column vector to 1 / 2 of the original; finally, through the fully connected layer l 2 neurons are connected to each node of the output layer, and the output S 2×1 dimensional medium distribution column vector.

[0043] An embodiment of the present invention further provides an electronic device, including:

[0044] one or more processors;

[0045] A memory for storing one or more programs;

[0046] When the one or more programs are executed by the one or more processors, the one or more processors implement the electrical capacitance tomography image reconstruction method in the embodiment.

[0047] The embodiment of the present invention further provides a storage medium on which a computer program is stored. When the program is executed by a processor, the electrical capacitance tomography image reconstruction method in the embodiment is implemented.

Claims

1. A method for reconstructing an electrical capacitance tomography image, characterized in that: include: The first-level neural network is used to perform nonlinear image reconstruction from capacitance vector to low-resolution medium distribution, where the input layer of the first-level neural network is an N×N-dimensional capacitance matrix obtained by rearranging the normalized capacitance vector, where N is the number of electrodes; the output of the first-level neural network is an S1×1-dimensional medium distribution column vector, where S1 is the number of pixel units of the low-resolution medium distribution image; A second-level neural network is used to perform super-resolution reconstruction of images from low-resolution medium distribution to high-resolution medium distribution; wherein the input of the second-level neural network is the output of the first-level neural network, and the output of the second-level neural network is an S2×1-dimensional medium distribution column vector, where S2 is the number of pixel units of the high-resolution medium distribution image; The first-level neural network is a first-level convolutional neural network, including an input layer, an upsampling layer, a convolution layer, a fully connected layer, and an output layer; the input layer is an N×N-dimensional capacitance matrix obtained by rearranging the normalized capacitance vector; after the input layer, the input capacitance matrix is ​​de-pooled and up-sampled to expand the data dimension; the convolution layer adopts two-dimensional convolution, sets m1 n1×n1-dimensional convolution kernels, and the convolution kernel step size is 1; the fully connected layer is located after the convolution layer, and the highly abstract features after the convolution are integrated through l1 neurons, and finally an S1×1-dimensional medium distribution column vector is output; The second-level neural network is a second-level convolutional neural network, including an input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer; the input layer is a reconstructed S1×1-dimensional low-resolution medium distribution column vector; the convolution layer adopts one-dimensional convolution, sets m2 n2×1-dimensional convolution kernels, and the convolution kernel step size is 1; the pooling layer adopts one-dimensional maximum pooling to shorten the length of the column vector to 1 / 2 of the original; finally, the l2 neurons of the fully connected layer are connected to each node of the output layer to output the S2×1-dimensional medium distribution column vector.

2. The electrical capacitance tomography image reconstruction method according to claim 1, characterized in that: A Dropout layer is added after the convolutional layer and the fully connected layer of the first-level convolutional neural network to randomly delete some neurons and their connections during training to enhance the generalization ability of the network.

3. The electrical capacitance tomography image reconstruction method according to claim 1, characterized in that: A Dropout layer is set after the convolutional layer and the fully connected layer of the second-level convolutional neural network.

4. A capacitance tomography image reconstruction model, characterized in that: include: The first-level neural network performs nonlinear image reconstruction from capacitance vector to low-resolution medium distribution, wherein the input layer of the first-level neural network is an N×N-dimensional capacitance matrix obtained by rearranging the normalized capacitance vector, where N is the number of electrodes; the output of the first-level neural network is an S1×1-dimensional medium distribution column vector, where S1 is the number of pixel units of the low-resolution medium distribution image; The second-level neural network performs super-resolution reconstruction of the image from the low-resolution medium distribution to the high-resolution medium distribution; wherein the input of the second-level neural network is the output of the first-level neural network, and the output of the second-level neural network is an S2×1-dimensional medium distribution column vector, where S2 is the number of pixel units of the high-resolution medium distribution image; The first-level neural network is a first-level convolutional neural network, including an input layer, an upsampling layer, a convolution layer, a fully connected layer, and an output layer; the input layer is an N×N-dimensional capacitance matrix obtained by rearranging the normalized capacitance vector; after the input layer, the input capacitance matrix is ​​de-pooled and up-sampled to expand the data dimension; the convolution layer adopts two-dimensional convolution, sets m1 n1×n1-dimensional convolution kernels, and the convolution kernel step size is 1; the fully connected layer is located after the convolution layer, and the highly abstract features after the convolution are integrated through l1 neurons, and finally an S1×1-dimensional medium distribution column vector is output; The second-level neural network is a second-level convolutional neural network, including an input layer, a convolution layer, a pooling layer, a fully connected layer and an output layer; the input layer is a reconstructed S1×1-dimensional low-resolution medium distribution column vector; the convolution layer adopts one-dimensional convolution, sets m2 n2×1-dimensional convolution kernels, and the convolution kernel step size is 1; the pooling layer adopts one-dimensional maximum pooling to shorten the length of the column vector to 1 / 2 of the original; finally, the l2 neurons of the fully connected layer are connected to each node of the output layer to output the S2×1-dimensional medium distribution column vector.

5. A device, characterized in that: include: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the capacitance tomography image reconstruction method according to any one of claims 1 to 3.

6. A storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the electrical capacitance tomography image reconstruction method according to any one of claims 1 to 3 is implemented.

Images