A Micro-Raman Automatic Analysis and Identification Method

Abstract

The present invention provides a micro-Raman automatic analysis and identification method, the method comprising: S101, moving the sample to be analyzed to the optimal focal plane and updating the background; S102, video stream frame acquisition; S103, image segmentation; S104, particle morphology analysis; S105, centering and focusing on the target particle; S106, spectrum acquisition. The micro-Raman automatic analysis and identification method proposed by the present invention can effectively find new particles through background modeling and parameter updating methods; the distance transformation algorithm of the local particle contour can accelerate the acquisition of the center position of the maximum inscribed circle; through the full process of automated analysis and identification, the traditional particle micro-Raman experimental analysis time is greatly shortened, and in-situ micro-Raman analysis can be realized under unattended conditions.

Description

Technical Field

[0001] The invention relates to the technical field of particle identification, and in particular to a micro-Raman automatic analysis and identification method. Background Art

[0002] Optical microscopy is a common method for detecting particles. It uses the principle of optical imaging to allow observers to directly observe particles for counting and identification. This method is simple to operate and low-cost. However, the resolution of optical microscopes is physically limited. For smaller particles, especially nanoparticles, it cannot provide sufficient resolution. In addition, human operation and visual judgment may introduce a certain degree of subjective error. Most importantly, it is impossible to identify the type of particles. Summary of the invention

[0003] In view of this, an object of the present invention is to provide a micro-Raman automatic analysis and identification method to solve or at least partially solve the above-mentioned problems existing in the prior art.

[0004] To achieve the above-mentioned object of the invention, the present invention provides a micro Raman automatic analysis and identification method, which comprises the following steps:

[0005] S101, moving the sample to be analyzed to the best focal plane and updating the background: acquiring a frame image from the camera video stream, and updating the background using the current frame image;

[0006] S102, video stream frame acquisition: the camera is framed once at a certain interval to obtain the latest frame image, the background image is subtracted from the latest frame image through image processing, and the background modeling method is used to analyze whether new particles appear in the latest frame image;

[0007] S103, image segmentation: using an image segmentation algorithm to segment the newly added particles appearing in the latest frame image, and extracting the outline size and spatial morphology information representation of the particles;

[0008] S104, particle morphology analysis: characterizing the spatial morphology information of the extracted particle contour, obtaining the current particle contour position, area and Feret diameter, and using a distance transformation algorithm to calculate the centroid position of the current particle contour;

[0009] S105, centering and focusing the target particle: converting the centroid coordinates to real space coordinates, calculating the parameter amount that the displacement module needs to move, controlling the movement of the displacement stage in the X-axis and Y-axis directions so that the newly added particle is in the center of the image, and then controlling the displacement stage to translate up and down in the Z-axis direction, finding the current best focal plane through the automatic focusing algorithm, and completing automatic focusing;

[0010] S106, spectrum acquisition: based on the current best focal plane, turn on the laser to collect the Raman spectrum, turn off the laser after the acquisition is completed, identify the type of particles based on the collected Raman spectrum, and store the data.

[0011] Furthermore, the background modeling method uses a Gaussian mixture model to model the grayscale distribution of background image pixels using multiple Gaussian distribution functions, and continuously updates the model parameters and uses the difference between the pixel grayscale value and the statistical characteristics of the background model to determine whether it is a foreground target. The probability density function of the Gaussian mixture model is expressed as:

[0012]

[0013] in, is the pixel gray value The probability density function of is the number of Gaussian distributions, is the sequence number of the Gaussian distribution, is the multivariate Gaussian distribution function, is the mean of the Gaussian distribution, is the covariance matrix of the Gaussian distribution.

[0014] Furthermore, the image segmentation algorithm includes dual-window adaptive threshold segmentation, which first determines the threshold for large-sized particles in the image using the local mean in the large window, and then determines the threshold for remaining small-sized particles using the local mean in the small window.

[0015] Furthermore, the distance transformation algorithm is used to find the center of the largest inscribed circle in the contour as the center of mass position of the contour, the particle area is calculated by converting the pixel size occupied by the particle, and the Feret diameter is calculated by the length and width of the minimum circumscribed rectangle of the particle.

[0016] Furthermore, the auto-focus algorithm determines the optimal focusing position by measuring the image clarity and according to the image clarity evaluation result. The image standard deviation can be used as a reference value for judging the image clarity, which is expressed as:

[0017]

[0018] in, is the width of the image, is the height of the image, is the column number of the image pixel position, is the row number of the image pixel location, For images The pixel value corresponding to the position, is the image mean.

[0019] Furthermore, before the particle type is identified based on the collected Raman spectra, the Raman spectra are normalized using a one-dimensional linear interpolation method to normalize the spectral data to 200-3200 cm -1 Range, 512 data points are evenly distributed. One-dimensional linear interpolation is based on the assumption that at any position between two known data points, the data value changes linearly with the position. The one-dimensional linear interpolation is expressed as:

[0020]

[0021] in, are the coordinates of the first known data point, are the coordinates of the second known data point, is the position for interpolation, For location At the same time, data normalization is used to eliminate the influence of collecting Raman spectra.

[0022] Furthermore, after the Raman spectrum data is standardized, it is input into a deep learning model for microplastic classification and identification, and the classification results are output using logistic regression. The Raman spectrum and the classification results are stored. The Raman spectrum is divided into a measured Raman spectrum T and a standard library Raman spectrum L. The deep learning model for microplastic classification and identification uses a training set to train the deep learning model.

[0023] Furthermore, in the deep learning model architecture for microplastic classification and identification, T and L are input, T and L are respectively extracted through two sub-networks, and then the outputs of the two sub-networks are subtracted, and the results are sent to the backbone network for upsampling. Finally, the upsampling result is expanded into a 512 neuron structure and sent to the fully connected layer, and the result outputs the score of the degree of matching of the two input spectra. The sub-network includes three layers of one-dimensional convolutional neural networks and four layers of residual networks, and the backbone network includes four layers of residual networks and one layer of one-dimensional convolutional neural network.

[0024] Furthermore, the training set is generated by a simulation method, and the steps of generating the training set are as follows:

[0025] (1) Select a random number as the number of spectral peaks;

[0026] (2) Randomly select a position as the spectral peak;

[0027] (3) Selecting a random peak height and peak width for each randomly selected position;

[0028] (4) Use a Gaussian function combined with a Hanning window function to model the spectral peak position, peak height, and peak width to obtain basic spectral data without any baseline or noise;

[0029] (5) randomly generate multiple anchor points, use cubic spline interpolation to generate random baselines, and combine the random baselines onto the basic spectral data;

[0030] (6) Add random Gaussian white noise to obtain virtually generated spectral data;

[0031] (7) Randomly combine the generated basic spectral data, i.e., the basic spectral data generated in step (4), with the spectral data to which random peak height changes, random baselines, and random white noise are added. If the two randomly combined data are of the same source, the label is set to 1, otherwise it is set to 0.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The micro-Raman automatic analysis and identification method proposed in the present invention can effectively find new particles through background modeling, and can effectively extract particles of different sizes through image segmentation algorithm, ensuring the accuracy of the number of extracted particles and particle contours. The distance transformation algorithm of the local particle contour can accelerate the acquisition of the center position of the maximum inscribed circle. The full-process automated analysis shortens the analysis time of traditional particle micro-Raman experiments and can realize in-situ micro-Raman analysis under unattended conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 A flow chart of micro-Raman automatic analysis and identification provided by an embodiment of the present invention;

[0036] Figure 2 A schematic diagram of newly added particle detection provided by an embodiment of the present invention;

[0037] Figure 3 A schematic diagram of a dual-window adaptive threshold segmentation process provided by an embodiment of the present invention;

[0038] Figure 4 A dual-window adaptive threshold segmentation result diagram provided by an embodiment of the present invention;

[0039] Figure 5 A diagram of a distance transformation algorithm provided by an embodiment of the present invention;

[0040] Figure 6 A diagram of the architecture of a deep learning model for microplastic classification and identification provided in an embodiment of the present invention;

[0041] Figure 7 A schematic diagram of a simulation-generated training set provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The enumerated embodiments are only used to explain the present invention and are not used to limit the scope of the present invention.

[0043] This embodiment provides a specific implementation method of the micro-Raman automatic analysis and identification method, and the method includes the following steps:

[0044] like Figure 1 As shown, a schematic diagram of a micro-Raman automatic analysis and identification process is provided, including:

[0045] S101, moving the sample to be analyzed to the optimal focal plane and updating the background, specifically including:

[0046] Control the translation stage to move to the optimal focal plane for the collection window that is manually measured in advance, take a frame image of the camera video stream, and update it with the current frame image as the background. Figure 2 A is a schematic diagram of a background image obtained during a certain detection.

[0047] S102, video stream frame acquisition, specifically including:

[0048] The camera takes video frames at regular intervals to obtain the latest frame image. Figure 2 B is a schematic diagram of the current frame image obtained. Until the image is processed, that is, the background image is subtracted from the latest frame image, the background modeling method is used to analyze whether new particles appear in the latest frame image, and when new particles appear using the dual-window adaptive threshold segmentation, the current background image is updated and the next process is entered. Figure 2 The black part shown in C is the newly added particles marked after detection.

[0049] Background modeling method: It can effectively determine whether there are new particles in the current field of view to be analyzed. The Gaussian mixture model uses multiple Gaussian distribution functions to model the grayscale distribution of background image pixels. By continuously updating the model parameters and using the difference between the pixel grayscale value and the statistical characteristics of the background model, it determines whether it is a foreground target. The probability density function of the GMM model (Gaussian mixture model) is expressed as:

[0050]

[0051] in, is the pixel gray value The probability density function of is the number of Gaussian distributions, is the sequence number of the Gaussian distribution, is the multivariate Gaussian distribution function, is the mean of the Gaussian distribution, is the covariance matrix of the Gaussian distribution.

[0052] S103, image segmentation, specifically including:

[0053] By segmenting the newly added particles in the latest frame image through the image segmentation algorithm, the contour size and spatial morphological information of the target sample can be effectively extracted. Threshold segmentation is one of the more typical segmentation methods. The expression of threshold segmentation is:

[0054]

[0055] in, is the gray value of each pixel before grayscale transformation, is the gray value of each pixel after grayscale transformation, is the threshold value used for image segmentation. For adaptive threshold segmentation, the local mean of pixels is used to calculate the threshold. The expression for adaptive threshold determination is:

[0056]

[0057] in, is the grayscale value of the pixel at position (x, y), To find the average function, is a constant that controls the threshold value. For the dual-window adaptive threshold segmentation, the threshold value is first determined for the large-sized particles in the image using the local mean value in the large window, and then the threshold value is determined for the remaining small-sized particles using the local mean value in the small window. Figure 3 The figure shows the effect of dual-window adaptive threshold. Figure 3 A is the result of large window adaptive threshold segmentation. Figure 3 B is the result of segmentation by small window adaptive threshold. Figure 3 C is Figure 3 A detects contours with an area greater than a certain threshold and superimposes them on Figure 3 The result obtained by B is Figure 3 D is Figure 3 A detected contours merged Figure 3 C The final segmentation result obtained by detecting the contour.

[0058] S104. Particle morphology analysis, including:

[0059] The extracted contour is used to characterize the spatial morphology of the particles, and the current particle contour position, area, and Feret diameter are obtained. The above information is stored after characterization.

[0060] Specifically, Figure 4 The figure shows the final segmentation effect. All contours are effectively marked, and the distance transformation algorithm is used to find the center of the largest inscribed circle in the contour as the center of mass of the contour (marked). The area of ​​the particle is calculated by converting the pixel size occupied by the particle, and the Feret diameter is calculated by the length and width of the minimum circumscribed rectangle of the particle.

[0061] The steps of the distance transformation algorithm are as follows: (1) for each particle (pixel is 1), extract it into a blank image (pixel is 0) of the same size as the original image, and use the rectangular area circumscribed by the outline as the ROI area;

[0062] (2) Use the forward template in the ROI area, such as Figure 5 As shown in A, the image is scanned from top to bottom and from left to right. If the pixel value corresponding to the center point 0 of the template is 0, it is skipped. If it is 1, the sum of the pixel values ​​corresponding to each element in the template is calculated, and the center pixel value is the minimum value of these five sum values.

[0063] (3) Then use the backward template in the ROI area, such as Figure 5 As shown in B, the image is scanned from bottom to top and from right to left. If the pixel value corresponding to the center point 0 of the template is 0, it is skipped. If it is 1, the sum of the pixel values ​​corresponding to each element in the template is calculated, and the center pixel value is the minimum value of these five sum values.

[0064] (4) Find the maximum pixel value in the ROI area. Add the position of the upper left corner of the ROI area in the original image to its maximum value coordinate offset to obtain the position of the center of the largest inscribed circle of the contour in the original image.

[0065] S105, centering and focusing on the target particles, specifically including:

[0066] The center of mass coordinates are converted into real space coordinates, and the parameter amount that the displacement module needs to move is calculated. The movement of the displacement stage in the X-axis and Y-axis directions is controlled so that the newly added particles are in the center of the image. The displacement stage is then controlled to translate up and down in the Z-axis direction. The image standard deviation is calculated as the image clarity evaluation index, and the golden section search method in the automatic focusing algorithm is used to find the optimal focal plane to complete the automatic focusing.

[0067] Autofocus algorithm: can accurately determine the current optimal focal plane position. This process involves two key steps: measuring the clarity of the image and determining the optimal focus position based on the clarity evaluation results. The image standard deviation can be used as a reference value for judging image clarity. The calculation formula is:

[0068]

[0069] in, is the width of the image, is the height of the image, is the column number of the image pixel position, is the row number of the image pixel location, For images The pixel value corresponding to the position, is the image mean. The calculation formula of the image mean is:

[0070]

[0071] From a mathematical point of view, the search for the best focus position is to find the extreme value of the function. The ideal image clarity evaluation curve is usually considered to be a convex function, and the global optimal solution exists and is unique. The golden section search method, also known as the 0.618 search method, is a method to find the extreme value by continuously narrowing the range of the maximum value of the single-peak function. Its core steps are as follows:

[0072] (1) Marking: Select an interval [L, R] as the starting interval for the search. The positions of the two sides of the interval are the left endpoint L and the right endpoint R respectively. The length of the interval is set to x , the image has the best clarity within this interval;

[0073] (2) Evaluation: Move to L + 0.382 x and L + 0.618 x The image clarity evaluation function is calculated at , if the image clarity evaluation function on the left is greater than that on the right, then the new search interval is determined to be [L, L + 0.618 x ], otherwise the new search interval is [R - 0.618 x , R];

[0074] (3) Accuracy judgment: If the length of the search interval is less than a certain accuracy threshold requirement, stop searching; otherwise, repeat step (2).

[0075] S106, spectrum collection, specifically including:

[0076] First, based on the current optimal focal plane, the laser is turned on to collect the Raman spectrum. After the collection is completed, the laser is turned off and the Raman spectrum is normalized using a one-dimensional linear interpolation method to normalize the spectral data to 200-3200 cm -1 Range, 512 data points are evenly distributed. One-dimensional linear interpolation is based on the assumption that at any position between two known data points, the data value changes linearly with the position. The one-dimensional linear interpolation is expressed as:

[0077]

[0078] in, are the coordinates of the first known data point, are the coordinates of the second known data point, is the position for interpolation, For location At the same time, data normalization can eliminate the effects caused by changes in laser light source power, integration time, light intensity attenuation, etc. when collecting Raman spectra, and significantly improve the model training effect and performance. The normalization conversion formula is as follows:

[0079]

[0080] in, is the peak value of the known data point, is the minimum value of the spectrum peak, is the maximum value of the spectral peak, is the normalized peak value.

[0081] Finally, the Raman spectrum data is standardized and input into the pre-trained deep learning model for microplastic classification and identification to identify the type of particles. The classification results are output using logistic regression, and the Raman spectrum and classification results are stored. The Raman spectrum is divided into a measured Raman spectrum T and a standard library Raman spectrum L. If there are still particles to be analyzed at this time, return to step S104 until all the newly added particles are analyzed, otherwise return to step S101 and control the translation stage to move to the optimal focal plane for the collection window that was manually measured in advance.

[0082] Specifically, the deep learning model architecture for microplastic classification and identification is as follows: Figure 6 As shown, input T and L, T and L are respectively extracted through two sub-networks, and then the outputs of the two sub-networks are subtracted, and the results are sent to the backbone network for upsampling. Finally, the upsampling result is expanded into a 512 neuron structure and sent to the fully connected layer, and the result outputs the score of the degree of matching of the two input spectra. The sub-network includes three layers of one-dimensional convolutional neural networks and four layers of residual networks, and the backbone network includes four layers of residual networks and one layer of one-dimensional convolutional neural network.

[0083] Specifically, the training set required by the deep learning model for microplastic classification and identification is generated by simulation method. Figure 7 The following is a schematic diagram of the generated simulated training set. Figure 7 A is a positive sample, which means that the input spectrum T and the input spectrum L are the spectrum of the same material. Figure 7 B is a negative sample, which means that the input spectrum T and the input spectrum L are spectra of different substances.

[0084] The steps to generate the training set are as follows:

[0085] (1) First, a random number x is selected from the range of 5-15 as the number of spectral peaks;

[0086] (2) Randomly select x positions within the range of 512 data points as the spectral peak positions;

[0087] (3) Select a random peak height and peak width for each position. The range of peak height is 0.025-1, and the range of peak width is 5-21cm -1 ;

[0088] (4) Use the Gaussian function combined with the Hanning window function to model the spectral peak position, spectral peak height and spectral peak width, and obtain basic spectral data without any baseline and noise. The expression of the Hanning window function is:

[0089]

[0090] in, is the Hanning window coefficient, is the index of the coefficient, ranging from 0 to The peak Gaussian function expression of the known data point is:

[0091]

[0092] in, is the spectral peak height, is a mathematical constant, Here is the horizontal axis of the spectrum, is the spectral peak position, is the spectral peak width, when = When the spectrum peak reaches its maximum value , Determines the degree of relaxation of the spectral peak;

[0093] (5) Randomly generate 2-7 anchor points, use cubic spline interpolation to generate a random baseline, and combine the random baseline with the basic spectral data, where the mixing ratio of the baseline to the basic spectral data ranges from 0.1 to 0.8;

[0094] (6) Add random Gaussian white noise, with the noise value fluctuating in the range of 0.001 to 0.01 times the intensity of the basic spectral data, to obtain virtually generated spectral data;

[0095] (7) The generated basic spectrum, i.e., the basic spectrum data generated in step (4), is randomly combined with the spectrum data to which random peak height changes, random baselines, and random white noise are added. If the two randomly combined data are of the same source, the label is set to 1, otherwise it is set to 0. 1 and 0 are used as labels so that the output of the microplastic classification and recognition deep learning model can be directly compared with the label, which is convenient for calculating the loss function.

[0096] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A micro-Raman automatic analysis and identification method, characterized in that: include: S101, moving the sample to be analyzed to the best focal plane and updating the background: acquiring a frame image from the camera video stream, and updating the background using the current frame image; S102, video stream frame acquisition: the camera is framed once at a certain interval to obtain the latest frame image, the background image is subtracted from the latest frame image through image processing, and the background modeling method is used to analyze whether new particles appear in the latest frame image; S103, image segmentation: using an image segmentation algorithm to segment the newly added particles appearing in the latest frame image, and extracting the outline size and spatial morphology information representation of the particles; S104, particle morphology analysis: characterizing the spatial morphology information of the extracted particle contour, obtaining the current particle contour position, area and Feret diameter, and using a distance transformation algorithm to calculate the centroid position of the current particle contour; S105, centering and focusing the target particle: converting the centroid coordinates to real space coordinates, calculating the parameter amount that the displacement module needs to move, controlling the movement of the displacement stage in the X-axis and Y-axis directions so that the newly added particle is in the center of the image, and then controlling the displacement stage to translate up and down in the Z-axis direction, finding the current best focal plane through the automatic focusing algorithm, and completing automatic focusing; S106, spectrum acquisition: based on the current optimal focal plane, turn on the laser to collect the Raman spectrum, turn off the laser after the acquisition is completed, identify the type of particles according to the collected Raman spectrum, and store the data. Before identifying the type of particles according to the collected Raman spectrum, use a one-dimensional linear interpolation method to standardize the Raman spectrum data. After the Raman spectrum data is standardized, it is input into the microplastic classification and recognition deep learning model. The microplastic classification and recognition deep learning model uses a training set to train the deep learning model. The training set is generated by a simulation method. The training set generation steps are as follows: (1) Select a random number as the number of spectral peaks; (2) Randomly select a position as the spectral peak; (3) Selecting a random peak height and peak width for each randomly selected position; (4) Use a Gaussian function combined with a Hanning window function to model the spectral peak position, peak height, and peak width to obtain basic spectral data without any baseline or noise; (5) randomly generate multiple anchor points, use cubic spline interpolation to generate random baselines, and combine the random baselines onto the basic spectral data; (6) Add random Gaussian white noise to obtain virtually generated spectral data; (7) Randomly combine the generated basic spectral data, i.e., the basic spectral data generated in step (4), with the spectral data to which random peak height changes, random baselines, and random white noise are added. If the two randomly combined data are of the same source, the label is set to 1, otherwise it is set to 0.

2. The micro-Raman automatic analysis and identification method according to claim 1, characterized in that: The background modeling method uses a Gaussian mixture model to model the grayscale distribution of background image pixels using multiple Gaussian distribution functions. By continuously updating the model parameters and using the difference between the pixel grayscale value and the statistical characteristics of the background model to determine whether it is a foreground target, the probability density function of the Gaussian mixture model is expressed as: in, is the pixel gray value The probability density function of is the number of Gaussian distributions, is the sequence number of the Gaussian distribution, is the multivariate Gaussian distribution function, is the mean of the Gaussian distribution, is the covariance matrix of the Gaussian distribution.

3. The micro-Raman automatic analysis and identification method according to claim 1, characterized in that: The image segmentation algorithm includes dual-window adaptive threshold segmentation, which first determines the threshold for large-sized particles in the image using the local mean in the large window, and then determines the threshold for the remaining small-sized particles using the local mean in the small window.

4. The micro-Raman automatic analysis and identification method according to claim 1, characterized in that: The distance transformation algorithm is used to find the center of the largest inscribed circle in the contour as the center of mass position of the contour, the particle area is calculated by converting the pixel size occupied by the particle, and the Feret diameter is calculated by the length and width of the minimum circumscribed rectangle of the particle.

5. The micro-Raman automatic analysis and identification method according to claim 1, characterized in that: The auto-focus algorithm determines the optimal focusing position by measuring the image clarity and according to the image clarity evaluation result. The image standard deviation can be used as a reference value for judging the image clarity, which is expressed as: in, is the width of the image, is the height of the image, is the column number of the image pixel position, is the row number of the image pixel location, For images The pixel value corresponding to the position, is the image mean.

6. The micro-Raman automatic analysis and identification method according to claim 1, characterized in that: The one-dimensional linear interpolation method is used to standardize the Raman spectrum data to 200-3200 cm -1 Range, 512 data points are evenly distributed. One-dimensional linear interpolation is based on the assumption that at any position between two known data points, the data value changes linearly with the position. The one-dimensional linear interpolation is expressed as: in, are the coordinates of the first known data point, are the coordinates of the second known data point, is the position for interpolation, For location At the same time, data normalization is used to eliminate the influence of collecting Raman spectra.

7. The micro-Raman automatic analysis and identification method according to claim 6, characterized in that: The microplastic classification and identification deep learning model uses logistic regression to output classification results, and stores Raman spectra and classification results. The Raman spectra are divided into measured Raman spectra T and standard library Raman spectra L.

8. The micro-Raman automatic analysis and identification method according to claim 7, characterized in that: In the deep learning model architecture for microplastic classification and identification, T and L are input, and T and L are respectively extracted through two sub-networks. Then the outputs of the two sub-networks are subtracted, and the results are sent to the backbone network for upsampling. Finally, the upsampling result is expanded into a 512 neuron structure and sent to the fully connected layer. The result outputs the score of the degree of matching of the two input spectra. The sub-network includes three layers of one-dimensional convolutional neural networks and four layers of residual networks, and the backbone network includes four layers of residual networks and one layer of one-dimensional convolutional neural network.

Images