A neural network-based laser-ultrasound imaging method

Abstract

This invention belongs to the field of laser ultrasound imaging technology, specifically disclosing a laser ultrasound imaging method based on neural networks, comprising the following steps performed sequentially: S1, acquiring the A-scan signal; S2, decomposing and reconstructing the A-scan signal through empirical mode decomposition; S3, extracting the temporal features of the ultrasound signal using pooling; S4, establishing a neural network and training it multiple times to obtain multiple neural networks with different representation capabilities; S5, inputting the temporal features of the ultrasound signal of the sample to be tested into each trained neural network, and filling the recognition results into the corresponding positions according to the detection point order to form a pixel matrix; S6, superimposing all pixel matrices, setting a threshold to evaluate the state of each detection point, and generating the corresponding pixel image. The method of this invention is insensitive to noise and can be used to process low signal-to-noise ratio signals, improving the accuracy of material state identification. This invention is applicable to detecting the state of materials.

Description

Technical Field

[0001] This invention belongs to the field of laser ultrasound imaging technology, specifically a laser ultrasound imaging method based on neural networks. Background Technology

[0002] Ultrasonic testing, with its advantages of being radiation-free and highly accurate, is one of the most widely used non-destructive testing methods currently available. Laser ultrasonic testing is a novel ultrasonic testing technology that utilizes the thermoelastic effect of a laser to excite ultrasonic waves and employs optical methods such as laser interference or the Doppler effect to detect these waves. Thermoelastic excitation occurs when the laser power density is less than the material's damage threshold, causing the laser to act on the material and raise its temperature. Since temperature decreases with increasing distance within the material, the rapid thermal expansion in the high-temperature region is constrained by the surrounding low-temperature, unexpanded material, thus generating thermal stress. This process is strictly non-destructive. Therefore, laser ultrasound possesses a series of advantages, including non-contact operation, high frequency, and a wide ultrasonic frequency range, and can be applied to the non-destructive testing of metallic or composite materials, making it a highly promising damage detection method. However, due to the influence of the testing method and material structure, laser ultrasonic testing is often accompanied by significant noise, leading to poor scanning imaging results. Currently, research on laser ultrasonic signal denoising is limited.

[0003] Neural networks are a programming paradigm that mimics the connection structure of biological neural networks. They enable computers to learn autonomously from observed data. Essentially, they are complex networks formed by the interconnection of numerous simple logical units. Data can be transferred between these units, giving the network a high degree of nonlinearity, thus allowing for complex logical operations and the fitting of nonlinear relationships. Neural networks can uncover deep information contained in signals and provide guidance for defect localization and classification. Furthermore, high-performance and robust neural networks can further improve detection efficiency, making their application in defect detection and classification a major future research direction. Summary of the Invention

[0004] The purpose of this invention is to provide a laser ultrasound imaging method based on neural networks to improve the accuracy of laser ultrasound detection and reduce imaging noise.

[0005] To achieve the above objectives, the present invention employs the following technical methods:

[0006] A laser ultrasound imaging method based on a neural network includes the following steps performed sequentially:

[0007] A laser ultrasound imaging method based on a neural network includes the following steps performed sequentially:

[0008] S1. Scan the sample using a laser ultrasonic system to obtain the A-scan signal;

[0009] S2. The A-scan signal is decomposed by empirical mode decomposition, the intrinsic mode functions are extracted, and the corresponding intrinsic mode functions are selected for reconstruction to obtain the reconstructed signal.

[0010] S3. Pool the reconstructed signal to extract the time-domain features of the ultrasound signal;

[0011] S4. Establish a neural network, input the time-domain features of the ultrasound signal into the neural network for multiple training sessions to obtain multiple neural networks with different representation capabilities;

[0012] S5. Input the time-domain features of the ultrasonic signals obtained after processing all detection points of the sample to be tested through steps S1-S3 into each trained neural network for recognition, and obtain the recognition result of each detection point of the sample to be tested. Fill the recognition results of all detection points in the same neural network into the corresponding positions according to the detection point order to form a pixel matrix for imaging.

[0013] S6. Overlay the pixel matrix formed by the recognition results of all neural networks, set a threshold to evaluate the state of each detection point, and generate the corresponding pixel image.

[0014] As a limitation, step S2 specifically includes:

[0015] S21. Find the maximum and minimum points of the original signal sequence signal(t). Fit the maximum points to the upper envelope and the minimum points to the lower envelope using the cubic spline interpolation function. Then, find the average value m(t) of the upper and lower envelopes.

[0016] S22. Subtract the average value m(t) from the original signal sequence signal(t) to obtain a new signal sequence h(t). Determine whether h(t) satisfies the condition that the difference between the number of extreme points and the number of zero crossing points is less than or equal to 1 and the average value of the upper envelope and the lower envelope is zero at any time. If not, replace the original signal sequence signal(t) with h(t) and repeat steps S21 and S22 until the condition is met. At this time, the new signal sequence that meets the condition is the first-order intrinsic mode function IMF1(t), and proceed to step S23.

[0017] S23. Subtract IMF1(t) from the original signal signal(t) to obtain the remainder r(t). Determine whether r(t) satisfies the conditions for a monotonic sequence or a constant sequence. If not, replace the original signal sequence signal(t) with r(t) and repeat steps S21-S23 until all intrinsic modulo functions are extracted. The extracted intrinsic modulo functions satisfy the formula:

[0018]

[0019] Where signal(t) is the original signal, and IMF i r(t) represents the i-th order intrinsic mode function, and r(t) represents the remainder term.

[0020] S24. Select the required intrinsic mode function for reconstruction to obtain the reconstructed signal, denoted as Restructure-signal(t).

[0021]

[0022] Where n1, n2, n3... are the orders of the intrinsic modulus functions used for reconstruction.

[0023] As a further clarification: Step S3 specifically includes:

[0024] S31. Pool the reconstructed signal Restructure-signal(t) according to the following formula:

[0025] Input-signal(t)=βdown(Restructure-signal(t))

[0026] Where Input-signal(t) represents the time-domain characteristics of the ultrasound signal, β is the amplification factor of the feature map, and down() is the downsampling function;

[0027] The length N of the sequence after pooling is:

[0028]

[0029] Where W is the length of the original sequence of the restructure-signal(t), F is the pooling kernel size, S is the stride, and P is the filling;

[0030] S32. The time-domain characteristic Input-signal(t) of the ultrasound signal is normalized according to the following formula:

[0031]

[0032] Wherein, [Input-signal(t)] nom This is the result after normalization.

[0033] As another limitation: step S4 specifically includes:

[0034] S41. Add state labels to the time-domain features of the ultrasound signal obtained in step S3, with damage state as 1 and normal state as 0, and divide them into training set and test set.

[0035] S42. Build a neural network, train the neural network multiple times using the training set, verify the trained neural network using the test set and output the accuracy and loss function, adjust the hyperparameters until the accuracy and loss function reach their optimal values, and obtain multiple neural networks with different representation capabilities.

[0036] As another limitation: step S6 specifically includes:

[0037] S61. The formula for superimposing the pixel matrix formed by the recognition results of all neural networks is:

[0038]

[0039] Among them, Picture new1 Picture is a matrix formed by superimposing the pixel matrices generated from the recognition results of all neural networks. j The pixel matrix formed by the recognition results of all detection points in the same neural network, where L is the number of neural networks with different representation capabilities;

[0040] S62. Set thresholds to evaluate the status of each detection point and determine the Picture. new1 The system checks whether the values ​​at all positions in the matrix exceed a set first threshold. If a value at a certain position exceeds the first threshold, the recognition result at that position is considered damaged and is set to 1; if a value at a certain position does not exceed the first threshold, the recognition result at that position is considered normal and is set to 0. The resulting values ​​are then used to fill the original positions to form a new pixel matrix (Picture). new2 ;

[0041] S63, For the pixel matrix Picture new2 The recognition result at position [x,y] is denoised in the pixel matrix Picture. new2 Select a [k×k] matrix with the [x,y] position as its core, and sum all the values ​​in the [k×k] matrix. The calculation formula is as follows:

[0042]

[0043] Where [k×k] represents a matrix with k rows and k columns, and Value is the sum of all values ​​in the [k×k] matrix. new2 [p,q] represents the pixel matrix Picture new2 The value in the p-th row and q-th column of the pixel matrix [x,y] represents the pixel matrix Picture. new2 The position of the x-th row and y-th column in the middle;

[0044] When the accumulated value exceeds the set second threshold, the pixel matrix Picture is considered to be... new2 The final identification result at position [x,y] is a damaged state, with a value of 1; when the accumulated value exceeds the set second threshold, the pixel matrix Picture is considered damaged. new2 The final recognition result at position [x,y] is in a normal state, with a value of 0;

[0045] S64, Picture new2 The recognition results of all positions are denoised according to step S63, and all the value results are filled back into the original position to form a denoised pixel matrix, Denoise-Picture.

[0046] The beneficial effects achieved by this invention, due to the adoption of the above-described solution, compared with the prior art, are as follows:

[0047] This invention provides a neural network-based laser ultrasound imaging method. It decomposes and reconstructs the A-scan signal using empirical mode decomposition to extract temporal features, and directly uses a neural network to identify the material state reflected by the signal. This overcomes the limitations of traditional imaging methods that use ultrasonic amplitude and phase for imaging, thus improving the accuracy of material state identification. The method is insensitive to noise and can be used to process low signal-to-noise ratio signals. Compared to traditional methods, it has a higher degree of automation and superior identification accuracy, effectively reducing missed or false detections caused by manual interpretation. Pooling is used to extract the temporal features of the ultrasonic signal, compressing the length of the sequence and reducing the computational load during identification. Training multiple neural networks and superimposing the identification results improves the accuracy of material state identification. Selecting a portion of the pixel matrix centered on a certain point and setting a second threshold for denoising further improves the accuracy of material state identification.

[0048] This invention is applicable to detecting the state of materials. Attached Figure Description

[0049] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0050] Figure 1 This is a flowchart of a laser ultrasound imaging method based on a neural network, according to Embodiment 1 of the present invention.

[0051] Figure 2 The image shows a comparison between the imaging results achieved using the conventional method and the imaging results achieved using Example 1.

[0052] Figure 3 The image shows a comparison between the imaging results achieved using the conventional method and the imaging results achieved using Example 2. Detailed Implementation

[0053] The present invention will be further described below with reference to the embodiments. However, those skilled in the art should understand that the present invention is not limited to the following embodiments. Any improvements and equivalent changes made based on the specific embodiments of the present invention are within the scope of protection of the claims of the present invention.

[0054] Example 1: A laser ultrasound imaging method based on neural networks

[0055] A laser ultrasound imaging method based on a neural network includes the following steps performed sequentially:

[0056] S1. The laser ultrasonic A-scan signal of the braided carbon fiber composite material is acquired using a laser ultrasonic system with a pulse reflection method. The sampling frequency is 1.1 GHz, the signal length is 10000, and the braided carbon fiber composite material is divided into a detection grid with 40×40 detection points in the X and Y directions. The scanning interval is 2 mm, and the detection area is 80×80 mm. The acquired A-scan signal is stored in the computer in the form of a spreadsheet.

[0057] S2. The A-scan signal is decomposed using empirical mode decomposition (EMD), intrinsic mode functions are extracted, and the corresponding intrinsic mode functions are selected for reconstruction to obtain the reconstructed signal; specifically including:

[0058] S21. Find the maximum and minimum points of the original signal sequence signal(t). Fit the maximum points to the upper envelope and the minimum points to the lower envelope using the cubic spline interpolation function. Then, find the average value m(t) of the upper and lower envelopes.

[0059] S22. Subtract the average value m(t) from the original signal sequence signal(t) to obtain a new signal sequence h(t). Determine whether h(t) satisfies the condition that the difference between the number of extreme points and the number of zero crossing points is less than or equal to 1 and the average value of the upper envelope and the lower envelope is zero at any time. If not, replace the original signal sequence signal(t) with h(t) and repeat steps S21 and S22 until the condition is met. At this time, the new signal sequence that meets the condition is the first-order intrinsic mode function IMF1(t), and proceed to step S23.

[0060] S23. Subtract IMF1(t) from the original signal signal(t) to obtain the remainder r(t). Determine whether r(t) satisfies the conditions for a monotonic sequence or a constant sequence. If not, replace the original signal sequence signal(t) with r(t) and repeat steps S21-S23 until all intrinsic modulo functions are extracted. The extracted intrinsic modulo functions satisfy the formula:

[0061]

[0062] Where signal(t) is the original signal, and IMF i r(t) is the i-th eigenmode function, and r(t) is the remainder term. In this embodiment, the A-scan signal at different detection points can be decomposed into 8 or 9 eigenmode functions, i.e., n = 8 or 9.

[0063] S24. Select the required intrinsic mode function for reconstruction to obtain the reconstructed signal, denoted as Restructure-signal(t).

[0064]

[0065] Where n1, n2, n3... are the orders of the intrinsic modulus functions used for reconstruction. In this embodiment, n1 = 3, n2 = 4, n3 = 5, n4 = 6, n5 = 7.

[0066] S3. Pool the reconstructed signal to extract the time-domain features of the ultrasound signal; specifically including:

[0067] S31. Pool the reconstructed signal Restructure-signal(t) according to the following formula:

[0068] Input-signal(t)=βdown(Restructure-signal(t))

[0069] Where Input-signal(t) represents the time-domain characteristics of the ultrasound signal, β is the amplification factor of the feature map, and down() is the downsampling function;

[0070] The length N of the sequence after pooling is:

[0071]

[0072] Where W is the original sequence length of the reconstructed signal Restructure-signal(t), F is the pooling kernel size, S is the stride, and P is the filling. In this embodiment, the pooling kernel length is 50, i.e., F = 50, the stride is 50, i.e., S = 50, and there is no filling, i.e., P = 0. Max pooling is used, i.e., downsampling retains the maximum value of the data in the pooling kernel. After pooling, the sequence length is shortened to 200.

[0073] S32. The time-domain characteristic Input-signal(t) of the ultrasound signal is normalized according to the following formula:

[0074]

[0075] Wherein, [Input-signal(t)] nom This is the normalized result, [Input-signal(t)] nomSave it to the computer in the form of a spreadsheet;

[0076] S4. Establish a neural network by inputting the time-domain features of the ultrasound signal into the neural network for five training iterations, resulting in five neural networks with different representation capabilities; specifically including:

[0077] S41. Add state labels to the time-domain features of the ultrasound signal obtained in step S3, with damage state as 1 and normal state as 0, and divide them into training set and test set.

[0078] S42. Establish an ANN neural network. The ANN neural network contains 200 input layer neurons, 25 hidden layer neurons, and 2 output layer neurons, which are the same length as the time-domain feature sequence of the input ultrasound signal. Train the ANN neural network five times using the training set, and verify the trained ANN neural network using the test set and output the loss function. Adjust the hyperparameters until the accuracy and loss function are optimal, resulting in five ANN neural networks with different representation capabilities. In this embodiment, the ANN neural network achieves optimal accuracy and loss function after 100 training rounds with a learning rate of 0.01.

[0079] S5. Input the time-domain features of the ultrasonic signals obtained after processing all the detection points of the woven carbon fiber composite material to be tested through steps S1-S3 into each trained ANN neural network to obtain the identification results of each detection point of the sample to be tested. Fill the identification results of all detection points in the same neural network into the corresponding positions according to the detection point order to form a pixel matrix for imaging.

[0080] S6. Overlay the pixel matrices formed by the recognition results of all neural networks, set a threshold to evaluate the state of each detection point, and generate the corresponding pixel image; specifically including:

[0081] S61. The formula for superimposing the pixel matrix formed by the recognition results of all neural networks is:

[0082]

[0083] Among them, Picture new1 Picture is a matrix formed by superimposing the pixel matrices generated from the recognition results of all neural networks. j The pixel matrix is ​​formed by the recognition results of all detection points in the same neural network, and L is the number of neural networks with different representation capabilities. In this embodiment, L = 5.

[0084] S62. Set thresholds to evaluate the status of each detection point and determine the Picture. new1The system checks whether the values ​​at all positions in the matrix exceed a set first threshold. If a value at a certain position exceeds the first threshold, the recognition result at that position is considered damaged and is set to 1; if a value at a certain position does not exceed the first threshold, the recognition result at that position is considered normal and is set to 0. The resulting values ​​are then used to fill the original positions to form a new pixel matrix (Picture). new2 The first threshold set in this embodiment is 3.

[0085] S63, For the pixel matrix Picture new2 The recognition result at position [x,y] is denoised in the pixel matrix Picture. new2 Select a [k×k] matrix with the [x,y] position as its core, and sum all the values ​​in the [k×k] matrix. The calculation formula is as follows:

[0086]

[0087] Where [k×k] represents a matrix with k rows and k columns, and Value is the sum of all values ​​in the [k×k] matrix. new2 [p,q] represents the pixel matrix Picture new2 The value in the p-th row and q-th column of the pixel matrix [x,y] represents the pixel matrix Picture. new2 The position of the x-th row and y-th column in the middle;

[0088] When the accumulated value exceeds the set second threshold, the pixel matrix Picture is considered to be... new2 The final identification result at position [x,y] is a damaged state, with a value of 1; when the accumulated value exceeds the set second threshold, the pixel matrix Picture is considered damaged. new2 The final recognition result at position [x,y] is in a normal state, with a value of 0; the second threshold set in this embodiment is 5, and k = 3;

[0089] S64, Picture new2 The recognition results of all positions are denoised according to step S63. All the value results are filled back into the original position to form a denoised pixel matrix Denoise-Picture. Then, the corresponding pixel image is generated and output according to the denoised pixel matrix Denoise-Picture.

[0090] A comparison of the imaging results using traditional methods and those using this embodiment: Figure 2 As shown, the upper part of the image represents the imaging effect using the traditional method, and the lower part represents the imaging effect using this embodiment. Darker colors in the image represent the normal state, and lighter colors represent the damaged state. Figure 2It can be seen that the imaging method in this embodiment is not sensitive to noise and can be used to process low signal-to-noise ratio signals. Compared with traditional methods, it has a higher degree of automation, better recognition accuracy, and effectively reduces missed detections or false detections caused by manual interpretation.

[0091] Example 2: A laser ultrasound imaging method based on neural networks

[0092] The main technical solution of this embodiment is basically the same as that of Embodiment 1. Features not explained in this embodiment are explained in Embodiment 1 and will not be repeated here. The difference between this embodiment and Embodiment 1 is as follows: S42, an LSTM neural network is established. The LSTM neural network includes a first layer with 200 memory cells and a second layer with 150 memory cells. The LSTM neural network is trained five times using a training set. The trained LSTM neural network is verified using a test set, and the loss function is output. The hyperparameters are adjusted until the loss function reaches its optimal value, resulting in five LSTM neural networks with different representation capabilities. In this embodiment, the loss function of the LSTM neural network reaches its optimal value after 90 rounds of training.

[0093] A comparison of the imaging results using traditional methods and those using this embodiment: Figure 3 As shown, the upper part of the image represents the imaging effect using the traditional method, and the lower part represents the imaging effect using this embodiment. Darker colors in the image represent the normal state, and lighter colors represent the damaged state. Figure 3 It can be seen that the imaging method in this embodiment is not sensitive to noise and can be used to process low signal-to-noise ratio signals. Compared with traditional methods, it has a higher degree of automation, better recognition accuracy, and effectively reduces missed detections or false detections caused by manual interpretation.

Claims

1. A laser ultrasound imaging method based on neural networks, characterized in that, This includes the following steps performed sequentially: S1. Scan the sample using a laser ultrasonic system to obtain the A-scan signal; S2. The A-scan signal is decomposed by empirical mode decomposition, the intrinsic mode functions are extracted, and the corresponding intrinsic mode functions are selected for reconstruction to obtain the reconstructed signal. S3. Pool the reconstructed signal to extract the time-domain features of the ultrasound signal; S4. Establish a neural network, input the time-domain features of the ultrasound signal into the neural network for multiple training sessions to obtain multiple neural networks with different representation capabilities; S5. Input the time-domain features of the ultrasonic signals obtained after processing all detection points of the sample to be tested through steps S1-S3 into each trained neural network for recognition, and obtain the recognition result of each detection point of the sample to be tested. Fill the recognition results of all detection points in the same neural network into the corresponding positions according to the detection point order to form a pixel matrix for imaging. S6. Overlay the pixel matrix formed by the recognition results of all neural networks, set a threshold to evaluate the state of each detection point, and generate the corresponding pixel image.

2. The laser ultrasound imaging method based on a neural network according to claim 1, characterized in that, Step S2 specifically includes: S21. Find the maximum and minimum points of the original signal sequence signal(t). Fit the maximum points to the upper envelope and the minimum points to the lower envelope using the cubic spline interpolation function. Then, find the average value m(t) of the upper and lower envelopes. S22. Subtract the average value m(t) from the original signal sequence signal(t) to obtain a new signal sequence h(t). Determine whether h(t) satisfies the condition that the difference between the number of extreme points and the number of zero crossing points is less than or equal to 1 and the average value of the upper envelope and the lower envelope is zero at any time. If not, replace the original signal sequence signal(t) with h(t) and repeat steps S21 and S22 until the condition is met. At this time, the new signal sequence that meets the condition is the first-order intrinsic mode function IMF1(t), and proceed to step S23. S23. Subtract IMF1(t) from the original signal signal(t) to obtain the remainder r(t). Determine whether r(t) satisfies the conditions for a monotonic sequence or a constant sequence. If not, replace the original signal sequence signal(t) with r(t) and repeat steps S21-S23 until all intrinsic modulo functions are extracted. The extracted intrinsic modulo functions satisfy the formula: Where signal(t) is the original signal, and IMF i r(t) is the i-th order intrinsic mode function, and r(t) is the remainder term; S24. Select the required intrinsic mode function for reconstruction to obtain the reconstructed signal, denoted as Restructure-signal(t). Where n1, n2, n3... are the orders of the intrinsic modulus functions used for reconstruction.

3. The laser ultrasound imaging method based on a neural network according to claim 2, characterized in that, Step S3 specifically includes: S31. Pool the reconstructed signal Restructure-signal(t) according to the following formula: Input-signal(t)=βdown(Restructure-signal(t)) Where Input-signal(t) represents the time-domain characteristics of the ultrasound signal, β is the amplification factor of the feature map, and down() is the downsampling function; The length N of the sequence after pooling is: Where W is the length of the original sequence of the restructure-signal(t), F is the pooling kernel size, S is the stride, and P is the filling; S32. The time-domain characteristic Input-signal(t) of the ultrasound signal is normalized according to the following formula: Wherein, [Input-signal(t)] nom This is the result after normalization.

4. A laser ultrasound imaging method based on a neural network according to any one of claims 1-3, characterized in that, Step S4 specifically includes: S41. Add state labels to the time-domain features of the ultrasound signal obtained in step S3, with damage state as 1 and normal state as 0, and divide them into training set and test set. S42. Build a neural network, train the neural network multiple times using the training set, verify the trained neural network using the test set and output the accuracy and loss function, adjust the hyperparameters until the accuracy and loss function reach their optimal values, and obtain multiple neural networks with different representation capabilities.

5. A laser ultrasound imaging method based on a neural network according to any one of claims 1-3, characterized in that, Step S6 specifically includes: S61. The formula for superimposing the pixel matrix formed by the recognition results of all neural networks is: Among them, Picture new1 Picture is a matrix formed by superimposing the pixel matrices generated from the recognition results of all neural networks. j The pixel matrix formed by the recognition results of all detection points in the same neural network, where L is the number of neural networks with different representation capabilities; S62. Set thresholds to evaluate the status of each detection point and determine the Picture. new1 The system checks whether the values ​​at all positions in the matrix exceed a set first threshold. If a value at a certain position exceeds the first threshold, the recognition result at that position is considered damaged and is set to 1; if a value at a certain position does not exceed the first threshold, the recognition result at that position is considered normal and is set to 0. The resulting values ​​are then used to fill the original positions to form a new pixel matrix (Picture). new2 ; S63, For the pixel matrix Picture new2 The recognition result at position [x,y] is denoised in the pixel matrix Picture. new2 Select a [k×k] matrix with the [x,y] position as its core, and sum all the values ​​in the [k×k] matrix. The calculation formula is as follows: Where [k×k] represents a matrix with k rows and k columns, and Value is the sum of all values ​​in the [k×k] matrix. new2 [p,q] represents the pixel matrix Picture new2 The value in the p-th row and q-th column of the pixel matrix [x,y] represents the pixel matrix Picture. new2 The position of the x-th row and y-th column in the middle; When the accumulated value exceeds the set second threshold, the pixel matrix Picture is considered to be... new2 The final identification result at position [x,y] is a damaged state, with a value of 1; if the accumulated value Value does not exceed the set second threshold, the pixel matrix Picture is considered damaged. new2 The final recognition result at position [x,y] is in a normal state, with a value of 0; S64, Picture new2 The recognition results of all positions are denoised according to step S63, and all the value results are filled back into the original position to form a denoised pixel matrix, Denoise-Picture.

6. The laser ultrasound imaging method based on a neural network according to claim 4, characterized in that, Step S6 specifically includes: S61. The formula for superimposing the pixel matrix formed by the recognition results of all neural networks is: Among them, Picture new1 Picture is a matrix formed by superimposing the pixel matrices generated from the recognition results of all neural networks. j The pixel matrix formed by the recognition results of all detection points in the same neural network, where L is the number of neural networks with different representation capabilities; S62. Set thresholds to evaluate the status of each detection point and determine the Picture. new1 The system checks whether the values ​​at all positions in the matrix exceed a set first threshold. If a value at a certain position exceeds the first threshold, the recognition result at that position is considered damaged and is set to 1; if a value at a certain position does not exceed the first threshold, the recognition result at that position is considered normal and is set to 0. The resulting values ​​are then used to fill the original positions to form a new pixel matrix (Picture). new2 ; S63, For the pixel matrix Picture new2 The recognition result at position [x,y] is denoised in the pixel matrix Picture. new2 Select a [k×k] matrix with the [x,y] position as its core, and sum all the values ​​in the [k×k] matrix. The calculation formula is as follows: Where [k×k] represents a matrix with k rows and k columns, and Value is the sum of all values ​​in the [k×k] matrix. new2 [p,q] represents the pixel matrix Picture new2 The value in the p-th row and q-th column of the pixel matrix [x,y] represents the pixel matrix Picture. new2 The position of the x-th row and y-th column in the middle; When the accumulated value exceeds the set second threshold, the pixel matrix Picture is considered to be... new2 The final identification result at position [x,y] is a damaged state, with a value of 1; if the accumulated value Value does not exceed the set second threshold, the pixel matrix Picture is considered damaged. new2 The final recognition result at position [x,y] is in a normal state, with a value of 0; S64, Picture new2 The recognition results of all positions are denoised according to step S63, and all the value results are filled back into the original position to form a denoised pixel matrix, Denoise-Picture.

Images