CN115937340A - Image reconstruction algorithm of capacitance tomography system based on lstm - Google Patents
Image reconstruction algorithm of capacitance tomography system based on lstm Download PDFInfo
- Publication number
- CN115937340A CN115937340A CN202211439354.7A CN202211439354A CN115937340A CN 115937340 A CN115937340 A CN 115937340A CN 202211439354 A CN202211439354 A CN 202211439354A CN 115937340 A CN115937340 A CN 115937340A
- Authority
- CN
- China
- Prior art keywords
- dimensional
- neural network
- vector
- image reconstruction
- lstm
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Image Processing (AREA)
Abstract
The invention relates to an image reconstruction method of an LSTM-based capacitance tomography system, which comprises the following steps: (1) And constructing a capacitance tomography system model, obtaining a sensitivity field of the capacitance tomography system through Comsol finite element simulation software, and obtaining capacitance vectors under the condition of distribution of object fields at different positions and different shapes. And (2) obtaining a reconstructed image by using a Landweber algorithm. (3) And determining time series samples according to the rows, and obtaining an image reconstruction result by using an LSTM neural network model. (4) And determining time series samples according to columns, and obtaining an image reconstruction result by using an LSTM neural network model. (5) And fusing the two image results to obtain a final image reconstruction result. The invention improves the quality of the reconstructed image by the method, and leads the distribution of the reconstructed medium to be closer to the real distribution.
Description
Technical Field
The invention relates to an image reconstruction method of an LSTM-based capacitance tomography system, belonging to the technical field of process tomography.
Background
The electric capacitance tomography technology is one of electric tomography, and is a process imaging technology for multi-phase flow parameter detection formed and developed in the middle and later period of the 80 th century. It is an important process tomography technology developed along with computer technology and sensor technology based on medical CT technology. The principle of ECT technology can be described as: if two substances with different dielectric constants are mixed together in a pipe or a container, when the composition of each substance and the distribution thereof are changed, the equivalent dielectric constant of the mixture is changed, so that the measured capacitance value is changed accordingly. By adopting the multi-electrode array type capacitive sensor, a plurality of capacitance measured values reflecting the dielectric constant distribution of the mixture can be provided by mutually combining the electrodes, and the capacitance measured values are taken as projection data, and an image reflecting the medium distribution condition of the pipeline or the device in a certain measured area can be reconstructed by adopting a certain image reconstruction algorithm. ECT is an intelligent real-time process parameter detection system which acquires two-dimensional and three-dimensional visual information of the section or space condition of a pipeline or a container in a complex industrial process in a non-invasive manner. This provides an effective way to detect some parameters whose characteristics are complicated and varied and which are difficult to detect by conventional methods. At present, a great deal of work is done on the ECT technology at home and abroad, the research content relates to the fields of chemical industry, petroleum, metallurgy, energy, power, light industry, nuclear energy and the like, and abundant scientific research achievements are obtained in the fields.
The ECT is not mature at present, and has the problems of undercharacterization, morbidity, nonlinearity, soft field performance and the like, so that the defects of low reconstructed image precision, non-ideal medium edge reconstruction effect and the like of the conventional image reconstruction algorithm are caused. Therefore, the method has important significance in researching the ECT image reconstruction algorithm.
Disclosure of Invention
The invention aims to solve the technical problem of an image reconstruction method of a capacitance tomography system based on an LSTM neural network, under the condition that capacitance projection data is unchanged, on the basis of reconstructing an image by a traditional method, a time sequence statement is constructed, and the image reconstruction task of capacitance tomography is completed by utilizing the long-time and short-time sequence memory capacity of the LSTM, so that the nonlinearity of the capacitance tomography system can be overcome to a certain extent. Therefore, the quality of the reconstructed image is improved, and the reconstructed medium distribution is closer to the real distribution.
In order to solve the problems, the specific technical scheme of the invention is as follows:
the image reconstruction method of the capacitance tomography system based on the LSTM comprises the following steps:
(1) Constructing a capacitance tomography system model, and obtaining a sensitivity field of a capacitance tomography system through Comsol finite element simulation software; and obtaining capacitance vectors under the condition of object field distribution at different positions and in different shapes.
(2) And (3) obtaining a reconstructed image by utilizing a Landweber algorithm according to the sensitivity field and the capacitance vector obtained in the step (1).
(3) And determining time series samples according to the rows, and obtaining an image reconstruction result by using an LSTM neural network model.
(3.1) converting the reconstructed images obtained in step (2) into an input "time series" with time steps: changing the reconstructed images obtained in the step (2) into one-dimensional vectors according to the head-to-tail connection of the lines; and rearranging the one-dimensional vector into an N-by-M-dimensional matrix, wherein the N rows can be regarded as an input time sequence with time steps, and each row of vector is a word embedding vector representing the word of the input sequence.
(3.2) converting the flow pattern image set by the Comsol finite element simulation software obtained in the step (1) into an output 'time sequence' with a time step: connecting the flow pattern images set by the Comsol finite element simulation software in the step (1) end to end according to rows to form a one-dimensional vector; the one-dimensional vector is rearranged into an N-M-dimensional matrix, the N rows can be regarded as an output time sequence with time steps, and each row of vectors is a word embedding vector representing the word of the output sequence.
And (3.3) training the neural network of the LSTM neural network by using the input sample set and the output sample set obtained in the steps (3.1) and (3.2), and storing the neural network model after training.
(3.4) in the inference stage, loading a neural network model, and changing head-to-tail connection of reconstructed images obtained by a Landweber algorithm into a one-dimensional vector according to rows; rearranging the one-dimensional vector into an N x M dimensional matrix, then sequentially sending the N row vectors into an LSTM neural network, and finally outputting the N x M dimensional matrix by the neural network; and then, the N-M dimensional matrix is connected according to the line head and the line tail and rearranged into a one-dimensional vector, and the vector can be rearranged into a two-dimensional square matrix N-N, which is the corresponding reconstructed image.
(4) And determining time series samples according to columns, and obtaining an image reconstruction result by using an LSTM neural network model.
(4.1) converting the reconstructed images obtained in step (2) into an input "time series" with time steps: changing the reconstructed images obtained in the step (2) into one-dimensional vectors according to the head-to-tail connection of columns; and rearranging the one-dimensional vector into an N-by-M-dimensional matrix, wherein the N rows can be regarded as an input time sequence with time steps, and each row of vector is a word embedding vector representing the word of the input sequence.
(4.2) converting the flow pattern image set by the Comsol finite element simulation software obtained in the step (1) into an output 'time sequence' with a time step: connecting the flow pattern images set by the Comsol finite element simulation software in the step (1) end to end according to a column to form a one-dimensional vector; the one-dimensional vector is rearranged into an N-by-M-dimensional matrix, the N rows can be regarded as an output time sequence with time steps, and each row of vectors is a word embedding vector representing the word of the output sequence.
And (4.3) training the neural network of the LSTM neural network by using the input sample set and the output sample set obtained in the steps (4.1) and (4.2), and storing the neural network model after training.
(4.4) in the inference stage, loading a neural network model, and changing the head-to-tail connection of reconstructed images obtained by the Landweber algorithm into one-dimensional vectors according to columns; rearranging the one-dimensional vector to form an N x M-dimensional matrix, then sequentially sending the N row vectors into an LSTM neural network, and finally outputting the N x M-dimensional matrix by the neural network; and rearranging the N x M dimensional matrix into a one-dimensional vector according to the row end to end, wherein the vector can be rearranged into a two-dimensional square matrix N x N, which is the corresponding reconstructed image.
(5) And fusing the results of the two images to obtain a final image reconstruction result: and (4) adding corresponding pixels of the image reconstruction results obtained in the step (3) and the step (4), and taking the average value as the pixel value of the fused image, so that the final image reconstruction result can be obtained.
The invention has the beneficial effects that:
the invention adopts an LSTM neural network structure, and can fully utilize the characteristic of nonlinear mapping of the neural network; and the characteristic that the LSTM neural network processes the sequence of the natural language is also utilized to change the reconstructed image into a time sequence statement with a time step, so that the relevance among image pixels is well utilized to a certain extent, and a coding library is not required to be established. Meanwhile, the invention provides two different coding schemes of time sequence and fuses the reconstructed images obtained by the two coding schemes, thus the invention can obviously improve the quality of the reconstructed images and ensure that the distribution of the reconstructed medium is closer to the real distribution. The invention also provides a new way and means for the research and application of capacitance tomography, and has good practical application value.
Drawings
FIG. 1 is a schematic diagram of the operation principle of the LSTM neural network.
Fig. 2 is a flow chart of an image reconstruction algorithm for an LSTM-based electrical capacitance tomography system.
Fig. 3 is a schematic diagram based on the image fusion principle of the LSTM neural network.
Fig. 4 is a graphical display example after image encoding.
Detailed Description
As shown in FIG. 1, the operation principle of the LSTM neural network is shown, and in the LSTM module of FIG. 1, the cell state at the last time t-1 is C t-1 The hidden layer state at the last time t-1 is h t-1 The input state at the current time t is X t Wherein h is t But also as an output at time t. Where X is t Is a current vector h after image coding t Is a current output vector of the LSTM neural network.
Fig. 2 shows a flow chart of an image reconstruction algorithm of the capacitance tomography system based on the LSTM neural network. The method mainly comprises the following steps: data preparation, LSTM neural network model training, data reasoning, image reconstruction, image fusion and the like. Wherein the data preparation mainly comprises obtaining a sample through Comsol, obtaining a reconstructed image result through Landweber, and changing the reconstructed image obtained by Landweber and Comsol set flow pattern into a time sequence sample according to a row rule or a column rule respectively. The LSTM neural network model training mainly comprises the steps of sample division, LSTM neural network structure parameters, neural network training and model storage. The data reasoning is mainly to carry out reasoning according to the model and input data to obtain an optimal result. The image reconstruction mainly utilizes the reasoning result to obtain the image reconstruction result required by the capacitance tomography system. The image fusion mainly comprises the step of fusing image results coded according to rows and columns to obtain the final image reconstruction result of the capacitance tomography system.
Fig. 3 is a schematic diagram based on the image fusion principle of the LSTM neural network. Firstly, determining time sequence samples according to rows and columns, obtaining corresponding image reconstruction results by using an LSTM neural network model, and then fusing the two images to obtain the final image reconstruction results of the capacitance tomography system.
An example of a graphical display after image encoding is shown in fig. 4. An original image (40 x 40) is divided into one-dimensional vectors, which are then reconstructed into a two-dimensional matrix of 4 rows of 400 columns each. Here, each line is coded into a two-dimensional image (20 × 20) for more intuitive display of the coding effect.
An image reconstruction method of a capacitance tomography system based on an LSTM neural network takes the example of determining time sequence samples according to rows, and comprises the following steps:
(1) Data preparation phase
1.1 Constructing a capacitance tomography system model through Comsol finite element simulation software, and obtaining a sensitivity field of the capacitance tomography system; the device is provided with a single flow pattern, two flow patterns, three flow patterns, four flow patterns, five flow patterns, circulation flow patterns, laminar flow patterns and the like, and the capacitance vectors under the distribution conditions of object fields of different positions and different shapes of the flow patterns are obtained as much as possible.
1.2 Based on the sensitivity field and the capacitance vector obtained in the step 1.1), utilizing Landweber algorithm to obtain reconstructed images of various set flow patterns.
1.3 The reconstructed images obtained in the step 1.2) are connected end to end according to 'lines' to form one-dimensional vectors; the one-dimensional vector is rearranged into an N-M-dimensional matrix, the N rows can be regarded as an input time sequence with time steps, and each row of vectors is a word embedding vector representing the word of the input sequence. Thus the reconstructed images obtained in step 1.2) become the input "time series" with time steps. Accordingly, the output "time series" with time steps can be obtained by processing the flow patterns set by the Comsol finite element simulation software in the same way.
(2) LSTM neural network model training phase
2.1 LSTM is a nn. LSTM module function in PyTorch framework, the number of each input and output "time series" word is 4, the number of input matrix features is 400, the number of hidden layer nodes is 500, the number of hidden layer layers is 3, and the output dimension is 400.
2.2 Determine the training samples, take 90% of the samples in the sample library as the training samples, and the rest samples as the testing samples.
2.3 Training the LSTM neural network, and after training is complete, saving the neural network model.
(3) Data reasoning phase
Loading a neural network model, and changing the head-to-tail connection of reconstructed images obtained by the Landweber algorithm into one-dimensional vectors according to the 'rows'; rearranging the one-dimensional vector to form an N-M-dimensional matrix, and then sequentially sending the N row vectors into an LSTM neural network to obtain an optimal result; and finally obtaining the matrix with the dimension of N x M. Here N =4,m =400.
(4) Image reconstruction phase
And (4) in the image reconstruction stage, data integration is mainly carried out by using the optimal reasoning result obtained in the step (3), so that an image reconstruction result required by the capacitance tomography system is obtained. That is, the N x M dimensional matrix is rearranged into a one-dimensional vector according to the row head-to-tail connection, and then the vector is rearranged into a two-dimensional square matrix N x N, which is the corresponding reconstructed image. Here N =4,m =400,n =40.
(5) Image reconstruction algorithm for LSTM neural networks that determine "time series" samples by "column
The data preparation stage for determining the time series samples according to the column is to process the encoding process of the step 1.3) according to the head-to-tail connection of the column, and other processes are not changed, so that another encoding result can be obtained. The neural network model training phase for determining "time series" samples by "column" is the same as step (2) above. In the data reasoning stage, the end-to-end connection of reconstructed images obtained by the Landweber algorithm according to a column is changed into a one-dimensional vector; rearranging the one-dimensional vector to form an N x M-dimensional matrix, and then sequentially sending the N row vectors into an LSTM neural network to obtain an optimal result. The image reconstruction phase is identical to step (4) above.
(6) Image fusion phase
The LSTM neural network is trained according to two different encoding modes of 'row' and 'column', and two different image reconstruction results can be obtained. And adding the pixels corresponding to the two image reconstruction results, and taking the average value as the pixel value of the fused image, thereby obtaining the final image reconstruction result.
The above description is only a preferred embodiment of the present invention, and the method of the present invention can also be applied to other tomography systems such as electromagnetic tomography and electrical resistance tomography. It should be noted that variations and modifications can be made by those skilled in the art without departing from the principle of the present invention, and they should also be considered as falling within the scope of the present invention.
Claims (4)
1. The image reconstruction method of the capacitance tomography system based on the LSTM is characterized by comprising the following steps:
(1) Constructing a capacitance tomography system model, and obtaining a sensitivity field of the capacitance tomography system through Comsol finite element simulation software; obtaining capacitance vectors under the condition of distribution of object fields at different positions and in different shapes;
(2) Obtaining a reconstructed image by utilizing a Landweber algorithm according to the sensitivity field and the capacitance vector obtained in the step (1);
(3) Determining time sequence samples according to the rows, and obtaining an image reconstruction result by using an LSTM neural network model;
(4) Determining time sequence samples according to the columns, and obtaining an image reconstruction result by using an LSTM neural network model;
(5) And fusing the two image results to obtain a final image reconstruction result.
2. The image reconstruction method of the LSTM-based electrical capacitance tomography system of claim 1, wherein in the step (3), the specific method is:
(3.1) converting the reconstructed images obtained in step (2) into an input "time series" with time steps: changing the reconstructed images obtained in the step (2) into one-dimensional vectors according to the head-to-tail connection of the lines; rearranging the one-dimensional vector into an N-M-dimensional matrix, wherein the N rows are regarded as input 'time sequence' with time steps, and each row of vector is a 'word embedding vector' representing the word of the input sequence;
(3.2) converting the flow pattern image set by the Comsol finite element simulation software obtained in the step (1) into an output 'time sequence' with a time step: connecting the flow pattern images set by the Comsol finite element simulation software in the step (1) end to end according to 'rows' to form one-dimensional vectors; rearranging the one-dimensional vector into an N x M-dimensional matrix, wherein N rows are regarded as an output 'time sequence' with time steps, and each row of vector is a 'word embedding vector' representing the word of the output sequence;
(3.3) training a neural network of the LSTM neural network by using the input sample set and the output sample set obtained in the steps (3.1) and (3.2), and storing a neural network model after training is finished;
(3.4) in the inference stage, loading a neural network model, and changing reconstructed images obtained by a Landweber algorithm into one-dimensional vectors according to line head-to-tail connection; rearranging the one-dimensional vector to form an N x M-dimensional matrix, then sequentially sending the N row vectors into an LSTM neural network, and finally outputting the N x M-dimensional matrix by the neural network; and then, the N-M dimensional matrix is connected according to the line head and the line tail and rearranged into a one-dimensional vector, and the vector can be rearranged into a two-dimensional square matrix N-N, which is the corresponding reconstructed image.
3. The image reconstruction method of the LSTM-based electrical capacitance tomography system of claim 1, wherein the specific method in step (4) is:
(4.1) converting the reconstructed images obtained in step (2) into an input "time series" with time steps: changing the reconstructed images obtained in the step (2) into one-dimensional vectors according to the head-to-tail connection of columns; rearranging the one-dimensional vector into an N-by-M-dimensional matrix, wherein the N rows can be regarded as an input time sequence with time steps, and each row of vector is a word embedding vector for representing the word of the input sequence;
(4.2) converting the flow pattern images set by the Comsol finite element simulation software obtained in the step (1) into an output 'time sequence' with time steps: connecting the flow pattern images set by the Comsol finite element simulation software in the step (1) end to end according to a row to form a one-dimensional vector; rearranging the one-dimensional vector into an N-M-dimensional matrix, wherein the N rows can be regarded as an output 'time sequence' with time steps, and each row of vector is a 'word embedding vector' representing the word of the output sequence;
(4.3) training a neural network of the LSTM neural network by using the input sample set and the output sample set obtained in the steps (4.1) and (4.2), and storing a neural network model after training is finished;
(4.4) in an inference stage, loading a neural network model, and changing the head-to-tail connection of reconstructed images obtained by the Landweber algorithm according to a column into a one-dimensional vector; rearranging the one-dimensional vector to form an N x M-dimensional matrix, then sequentially sending the N row vectors into an LSTM neural network, and finally outputting the N x M-dimensional matrix by the neural network; and rearranging the N x M dimensional matrix into a one-dimensional vector according to the row end to end, wherein the vector can be rearranged into a two-dimensional square matrix N x N, which is the corresponding reconstructed image.
4. The image reconstruction method of the LSTM-based electrical capacitance tomography system of claim 1, wherein the step (5) is performed by: and (4) adding corresponding pixels of the image reconstruction results obtained in the step (3) and the step (4), and taking the average value as the pixel value of the fused image, so that the final image reconstruction result can be obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211439354.7A CN115937340A (en) | 2022-11-17 | 2022-11-17 | Image reconstruction algorithm of capacitance tomography system based on lstm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211439354.7A CN115937340A (en) | 2022-11-17 | 2022-11-17 | Image reconstruction algorithm of capacitance tomography system based on lstm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115937340A true CN115937340A (en) | 2023-04-07 |
Family
ID=86654982
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211439354.7A Pending CN115937340A (en) | 2022-11-17 | 2022-11-17 | Image reconstruction algorithm of capacitance tomography system based on lstm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115937340A (en) |
-
2022
- 2022-11-17 CN CN202211439354.7A patent/CN115937340A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111510157B (en) | Quantum error correction decoding method, device and chip based on neural network | |
| Bach et al. | Towards unambiguous edge bundling: Investigating confluent drawings for network visualization | |
| CN112270406B (en) | Nerve information visualization method of brain-like computer operating system | |
| CN110674742B (en) | Remote sensing image road extraction method based on DLinkNet | |
| CN113313180B (en) | Remote sensing image semantic segmentation method based on deep confrontation learning | |
| CN110580727B (en) | Depth V-shaped dense network imaging method with increased information flow and gradient flow | |
| CN115545166A (en) | Improved ConvNeXt convolutional neural network and remote sensing image classification method thereof | |
| CN115641263A (en) | Single-power equipment infrared image super-resolution reconstruction method based on deep learning | |
| CN115937340A (en) | Image reconstruction algorithm of capacitance tomography system based on lstm | |
| CN108769674A (en) | A kind of video estimation method based on adaptive stratification motion modeling | |
| CN115762675A (en) | Method and system for predicting mechanical property of truss type lattice structure | |
| CN110020379B (en) | Link prediction method based on deep dynamic network embedded representation model | |
| Guo et al. | Speedy and accurate image super‐resolution via deeply recursive CNN with skip connection and network in network | |
| CN113191947B (en) | Image super-resolution method and system | |
| CN116266377A (en) | Three-dimensional reconstruction method for microstructure of core image based on LSTM | |
| Zhang et al. | HybridDenseU-Net: learning a multi-scale convolution and dense connectivity CNN for inverse imaging problems | |
| CN114119862A (en) | Garbage yard three-dimensional reconstruction method and system based on wide baseline light field imaging technology | |
| CN112150566A (en) | Dense residual error network image compressed sensing reconstruction method based on feature fusion | |
| CN111444614A (en) | Flow field reconstruction method based on graph convolution | |
| CN118135045A (en) | Electrical impedance tomography system error modeling method and device | |
| CN115861762B (en) | Plug-and-play infinite deformation fusion feature extraction method and application thereof | |
| KR102572429B1 (en) | Method, apparatus and storage for storing a program for multi-demensional matrix multiplication | |
| CN113033034B (en) | Mixed order vector finite element electromagnetic field calculation method and system | |
| CN115830160A (en) | Image reconstruction method of capacitance tomography system based on Transformer | |
| CN117723895B (en) | Distribution network fault section positioning method and device based on multi-terminal amplitude ratio matrix |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |