A high-precision ultrasonic imaging recognition method based on cosmic muon image

Abstract

The application discloses a kind of high-precision ultrasonic imaging identification methods based on universe μ picture, steps are as follows: S1, generate universe μ imaging feature map;S2, construct and train core defect detection network, output the defect probability map of the defect area of universe μ imaging feature map is identified;S3, the defect area of imaging object is ultrasonically imaged, and color mapping reconstruction is obtained ultrasonic color image;S4, respectively acquire the non-defect area mask of universe μ imaging feature map and the defect area mask of ultrasonic color image, and image fusion is obtained fusion image by;The method is suitable for large-scale high-precision nondestructive imaging detection, which can be realized on the basis of large-scale, no dose transmission imaging using universe μ, effectively identify the defect area and be fused with the ultrasonic imaging of defect area, realize the high-precision imaging of local defect position, and have the advantages of nondestructive imaging and completely no radiation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of non-destructive testing image recognition and processing technology, in particular to a high-precision ultrasonic imaging recognition method based on cosmic muon image. BACKGROUND

[0002] As a unique natural particle, cosmic muon has super strong penetration ability and can easily penetrate through complex medium layers in a closed space. This characteristic makes it have great potential in the detection application of closed space and complex medium scenes. However, current muon imaging schemes mostly rely on long-time cumulative detection and data processing process, which greatly limits the imaging speed and accuracy of cosmic muon imaging technology in real-time monitoring field.

[0003] In the field of non-destructive imaging, ultrasonic non-destructive testing technology can accurately detect cracks, pores and inclusions and other defects inside the material without damaging the material structure by using high-frequency ultrasonic pulses. Compared with X-ray imaging technology, it is completely free of radiation and safe and environmentally friendly.

[0004] However, ultrasonic non-destructive testing technology also faces many difficulties in the actual application of imaging objects in closed blocks. First, when ultrasonic waves propagate inside the object to be measured with complex structure, they will encounter multiple interfaces and defects, causing multi-path reflection and scattering phenomena, which makes the received signal complex and difficult to accurately distinguish and analyze, thereby affecting the accuracy of defect detection. Secondly, the area of the ultrasonic probe is relatively small, and in the actual detection process, the probe generally needs to be moved frequently, which limits the coverage and detection efficiency and makes it difficult to meet the needs of large-scale, high-precision non-destructive imaging detection. In addition, ultrasonic detection images can only provide relative position information, lacking absolute position information, which makes it difficult to overcome the matching and positioning problems when it is fused with other imaging technologies such as cosmic muon image. Therefore, how to effectively fuse the two imaging technologies to detect defects more accurately for complex structures has become a difficult problem to be solved. SUMMARY

[0005] The purpose of the present application is to provide a high-precision ultrasonic imaging recognition method based on cosmic muon image to solve the above technical problems.

[0006] To this end, the technical scheme of the present application is as follows:

[0007] A high-precision ultrasonic imaging recognition method based on cosmic muon image, the steps are as follows:

[0008] S1, lay two liquid scintillator muon detectors on the upper and lower paths of the closed block in which the imaging object is placed, collect the real track information of each muon entering and exiting the closed block, and generate a cosmic muon imaging feature map;

[0009] S2, construct and train a core defect detection network to identify the defect area of the cosmic muon imaging feature map generated in step S1, and output a defect probability map to determine the defect area of the imaging object;

[0010] S3, according to the recognition result of step S2, perform ultrasonic imaging on the defect area of the imaging object, then normalize the echo amplitude in the ultrasonic imaging to the interval [0, 255] and perform color mapping to obtain an ultrasonic color image that can be fused with the cosmic muon imaging feature map;

[0011] S4, obtain the non-defect area mask of the cosmic muon imaging feature map and the defect area mask of the ultrasonic color image respectively, and obtain a fusion image that clearly shows the defect position and shape of the imaging object in the closed block through image fusion.

[0012] Further, the specific implementation steps of step S1 are as follows:

[0013] S101, obtain the real muon scattering angle data of the closed block according to the real track information of each muon entering and exiting the closed block;

[0014] S102, construct a forward model to obtain the theoretical muon scattering angle data of the closed block by simulating the muon flux distribution;

[0015] S103, calculate the optimal solution of the density distribution of the closed block by the MLSD inversion algorithm;

[0016] S104, project the optimal solution of the density distribution of the closed block onto a two-dimensional plane and assign color values to generate a cosmic muon imaging feature map.

[0017] Further, step S102 is implemented by using Geant4 software; wherein the modeling model of the closed block and the imaging object is consistent with the actual situation, and the model of the object to be imaged is set to have a uniform theoretical density value, and the remaining space in the closed block is set to be full of air; one liquid scintillator muon detector is arranged in the middle of the upper path and the lower path of the closed block model, so that the scattering data collection area completely covers the closed block model, and the direction of cosmic muon emission is set to be random.

[0018] Further, the specific processing steps of step S103 are as follows:

[0019] S1031, divide the closed block 1 into a voxel grid, each voxel is assigned an initial density value;

[0020] S1032, define the objective function F: , wherein, △ θ i is the actual scattering angle of the first i μ, △ θ' i is the theoretical scattering angle of the first i μ, and N is the total number of μ;

[0021] S1033, using the least square method, taking minimizing the objective function F as the optimization goal, and continuously iterating to obtain the optimal density distribution result of the closed block.

[0022] Further, the specific implementation steps of step S2 are:

[0023] S201, construct a core defect detection network, the encoder of the network is composed of a first convolutional layer, a second convolutional layer, a third convolutional layer and a fourth convolutional layer connected in turn; each convolutional layer adopts a 3*3 convolutional layer, and the step length is 2; the decoder is composed of a first deconvolutional layer, a second deconvolutional layer, a third deconvolutional layer and a fourth deconvolutional layer; each deconvolutional layer adopts a 4*4 convolutional layer, and the step length is 2; wherein, the output end of the first convolutional layer is connected with the output end of the third deconvolutional layer and the input end of the first splicing module, the output end of the first splicing module is connected with the input end of the fourth deconvolutional layer; the output end of the second convolutional layer and the output end of the second deconvolutional layer are connected with the input end of the second splicing module, the output end of the second splicing module is connected with the input end of the third deconvolutional layer; the output end of the third convolutional layer and the output end of the first deconvolutional layer are connected with the input end of the third splicing module, and the output end of the third splicing module is connected with the input end of the second deconvolutional layer;

[0024] S202, construct a cosmic μ image data set for training the core defect detection network, which adopts the same method as step S1, respectively acquires a plurality of cosmic μ imaging feature maps of the imaging object in the closed block without defects, and a plurality of cosmic μ imaging feature maps of the imaging object with known core defects of different shapes and different types, adjusts the resolution uniformly, and forms a cosmic μ imaging feature map by pixel-level defect labeling and defect class label assignment;

[0025] S203, train the core defect detection network using the cosmic μ image data set;

[0026] S204, after adjusting the cosmic μ imaging feature map obtained by step S1 into an image with the same resolution, input it into the trained core defect detection network, and output a defect probability map.

[0027] Further, in the training process in step S203, the Adam optimizer is adopted, the initial learning rate is set to 1e -4 , the decay strategy is dynamic decay; the number of iterations is set to 50 rounds; the loss function L Combining Dice Loss and BCE Loss to alleviate the class imbalance problem, the expression is:

[0028] ,

[0029] In the formula, L Dice is the Dice loss function, L BCE is the BCE loss function, λ is the weight coefficient, and its value is .

[0030] Further, in step S3, ultrasonic imaging is realized by an ultrasonic imaging detector, which is arranged in the front or back of the closed block; wherein,

[0031] According to the thickness of the imaging object in the detection direction, a probe is selected; when the thickness of the imaging object in the detection direction is ≤20mm, it is considered as a thin object, and a high-frequency probe is used; when the thickness of the imaging object in the detection direction is >20mm, it is considered as a thick object, and a low-frequency probe is used;

[0032] The ultrasonic parameters suitable for ultrasonic imaging of the imaging object in the closed block are set, including:

[0033] 1) Gain, which is initially set to 40dB and is fine-tuned within ±20dB in subsequent ultrasonic detection;

[0034] 2) Detection frequency, which is set to 5MHz~10MHz for thin objects and 2MHz~4MHz for thick objects;

[0035] 3) Depth range, which is set to increase by 10mm~15mm in the thickness of the imaging object in the detection direction;

[0036] 4) Scanning speed, which is set to 10mm / s~50mm / s;

[0037] 5) Gate, which is set to the starting position of the detection in the closed space, the detection width is the width of the defect area, and the detection height is 2mm~10mm.

[0038] Further, the image reconstruction step of step S3 is:

[0039] 1) Select the echo amplitude of the ultrasonic image as the characteristic parameter, and normalize it to the numerical range interval of 0~255;

[0040] 2) For the normalized numerical range interval, set 0~30 as dark blue and represent large defects, set 31~80 as light blue and represent small defects, set 81~200 as green and represent normal regions, and set 201~255 as red and represent strong reflection interfaces; for each color, set the depth threshold at 10mm intervals, so that the ultrasonic pixel increases in depth with the increase of the quantization value;

[0041] 3) Based on the normalized data of step 1) and the color mapping settings of step 2), the image reconstruction algorithm based on the coherence factor is used to complete the image reconstruction;

[0042] 4) Use the Sobel edge enhancement algorithm to post-process the color image reconstructed by step 3).

[0043] Further, in step S4, the specific method for obtaining the non-defect area mask of the primordial muon imaging feature map is:

[0044] 1) Convert the defect probability map output by step S2 into a binary mask;

[0045] 2) Multiply the value of each pixel in the binary mask by 255, and then convert it to uint8 type;

[0046] 3) Define a 3x3 rectangular kernel for morphological closing operation, and define the closing operation expression;

[0047] 4) Use the method defined in step 3) to perform closing operation on the mask obtained in step 2), and set the iteration number of the closing operation to 2 times;

[0048] 5) Divide the value of each pixel in the mask obtained in step 4) by 255 to output a new binary mask;

[0049] 6) Perform a mask inversion operation on the new binary mask obtained in step 5) to obtain a color mask of the non-defect area of the primordial muon imaging feature map.

[0050] Further, in step S4, the specific steps for obtaining the defect area mask of the ultrasonic color image and performing image fusion with the non-defect area mask of the primordial muon imaging feature map are:

[0051] 1) Adjust the ultrasonic color image to have the same resolution as the primordial muon imaging feature map;

[0052] 2) Use the mask extraction module to extract the defect area mask in the ultrasonic color image;

[0053] 3) fusing the non-defect area mask obtained in step 1) and the defect area mask obtained in step 2); the weighted mixing formula of the fusion process is:

[0054] I fused = I RGB非缺陷 ×(1- α ) + I US缺陷 × α ,

[0055] In the formula, I RGB非缺陷 is the intensity of the non-defect area of the RGB image; I US缺陷 is the intensity of the defect area of the ultrasonic image; and α is a weight, and the value range of α is 0.4-0.8.

[0056] Compared with the prior art, the high-precision ultrasonic imaging recognition method based on the muon image of the universe is suitable for large-range high-precision non-destructive imaging detection, and on the basis of realizing large-range, dose-free transmission imaging by using the muon of the universe, the core defect detection network is constructed and trained to realize accurate recognition of the defect area on the muon imaging feature map of the universe, cooperate with the ultrasonic imaging of the defect area, and fuse with the muon imaging feature map of the universe to obtain a detection image with high-precision imaging of the local defect position of the imaged object, and has the advantages of non-destructive imaging and complete non-radiation. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is a flowchart of the high-precision ultrasonic imaging recognition method based on the muon image of the universe of the present application;

[0058] Figure 2 is a schematic diagram of the liquid scintillator muon detector laid out in two ways of upper and lower outside the closed block in which the imaging object is built in step S101 of the embodiment of the present application;

[0059] Figure 3 is a schematic diagram of the forward model constructed by using the Geant4 software in step S102 of the embodiment of the present application;

[0060] Figure 4 is the muon imaging feature map generated in step S104 of the embodiment of the present application;

[0061] Figure 5 is a structural schematic diagram of the core defect detection network in step S2 of the embodiment of the present application;

[0062] Figure 6A schematic diagram of red frame identification of a defect region identified by the core defect detection network in step S2 of the embodiment of the present application in the cosmic muon imaging feature map;

[0063] Figure 7 A schematic diagram of the setting position of the ultrasonic imaging detector in step S3 of the embodiment of the present application;

[0064] Figure 8 A schematic diagram of the initial ultrasonic imaging obtained in step S3 of the embodiment of the present application;

[0065] Figure 9 A schematic diagram of the reconstructed color ultrasonic imaging obtained in step S3 of the embodiment of the present application;

[0066] Figure 10 A schematic diagram of the fusion image obtained in step S4 of the embodiment of the present application. DETAILED DESCRIPTION

[0067] The present application will be further described below in conjunction with the accompanying drawings and specific embodiments, but the following embodiments are by no means any limitation on the present application.

[0068] Referring to Figure 1 The specific processing steps of the high-precision ultrasonic imaging recognition method based on the cosmic muon image are described as follows.

[0069] S1, obtaining a cosmic muon imaging feature map of an imaging object in a closed block.

[0070] The specific implementation steps of this step S1 are described as follows.

[0071] S101, laying a liquid scintillator muon detector 3 on the upper and lower two paths of the closed block 1 with the imaging object 2 built-in, so as to collect muon scattering data of the closed block; referring to Figure 2 In the actual scene setting, the imaging object 2 is a lead alloy block with uniform density distribution, wherein there is an irregular E-shaped crack in the B part of the imaging object 2, that is, the defect recognition target to be realized by non-destructive imaging in this embodiment.

[0072] In this step S101, in order to meet the imaging needs of the closed space, the liquid scintillator μ probe is laid along the 3 upper and lower two ways to obtain the scattering data of μ before and after passing through the closed block 1 respectively; at the same time, in order to ensure the continuity of the scattering imaging data on the imaging result, the liquid scintillator μ probe 3 located in the upper and lower ways is arranged in a linear and equidistant manner with multiple (3 in this embodiment), and the interval between the adjacent two liquid scintillator μ probes 3 located in the same way is determined according to the detection range of the liquid scintillator μ probe 3, so as to ensure that the scattering data acquisition area has a certain spatial overlap. Since the μ scattering imaging is realized by accumulating μ scattering data, therefore, in this step, the μ scattering data acquisition time is set to 2h.

[0073] In this step S101, the μ scattering data to be obtained is specifically: the real track information of each μ entering the closed block 1 X i and the real track information of each μ out of the closed block 1 X i ’ , i is the number of μ; and further, the angle change caused by scattering during the passage of each μ through the closed block 1, i.e. the real scattering angle △ θ i .

[0074] S102, constructing a forward model to obtain the theoretical μ scattering data of the closed block by simulating the μ flux distribution.

[0075] In this step S102, the interaction between cosmic μ and the imaging object 2 is simulated by the constructed forward model, so as to obtain the theoretical μ scattering data, which is used as a data basis for further comparison with the real μ scattering data measured in step S101 and solving to obtain the actual density distribution data of the imaging object 2.

[0076] In this embodiment, the forward model construction and simulation are realized by using Geant4 software; specifically, the forward model construction includes: first, modeling the closed block 1 and the imaging object 2 located in the closed block 1 in the software, and keeping the geometric models of the two consistent with the actual scene state in step S1; specifically, the closed block model is a hollow shell model, and the imaging object model is a solid model composed of a plurality of voxels; then, according to the properties of the lead alloy block of the real imaging object, the imaging object model is set to have a uniform theoretical lead alloy density value, and the remaining space inside the closed block model is set to be full of air; finally, one liquid scintillator μ probe is arranged in the middle of the upper and lower ways of the closed block model, and it is ensured that the scattering data acquisition area completely covers the closed block model, and the direction of the cosmic μ is set to be random; as shown in the following figure:Figure 3 Shown is a schematic diagram of the forward model constructed using Geant4 software (top view).

[0077] Then, Geant4 software can simulate the theoretical muon flux distribution based on the above model construction conditions, and then obtain theoretical muon flux data; specifically, the same as step S101, the theoretical muon flux data is: the theoretical track information of each muon penetrating the closed block 1 Y i and theoretical track information of passing through closed block 1 Y i ’ ,in, i is the number of the muon; further calculations yield the following: the theoretical angle change caused by scattering when each muon passes through the closed block, i.e. the theoretical scattering angle △ θ' i .

[0078] S103, calculating the optimal solution of the density distribution of the imaging object by using the MLSD inversion algorithm.

[0079] The specific implementation steps of step S103 are:

[0080] S1031 , dividing the closed block 1 into a voxel grid, and assigning an initial density value to each voxel; wherein the division accuracy of the voxel grid is determined according to actual needs to meet the resolution requirement of two-dimensional imaging.

[0081] In this embodiment, the initial density distribution of each voxel is consistent with the theoretical density value set in the forward model in step S102.

[0082] S1032. Define the objective function F as the sum of the squares of the differences between the actual scattering angles of each muon and the theoretical scattering angles. Then, use the correlation between the muon flux distribution and the closed block density distribution to deduce the optimal solution for the closed block density distribution.

[0083] Specifically, the expression of the objective function F is:

[0084] ,

[0085] Where, θ i For the i The actual scattering angle of a muon, △ θ' i For the i The theoretical scattering angle of muons, N is the total number of muons.

[0086] S1033, adopt the least square method, with the minimum objective function F as the optimization objective, and obtain the optimal density distribution result of the closed block through continuous iteration calculation.

[0087] In this embodiment, the density distribution range of the closed block obtained through step S103 is 1.1 kg / m 3 (about air density)~8.4×10 3 kg / m 3 , which is consistent with the actual scene.

[0088] S104, color the density distribution of the closed block obtained by step S103, and project it onto a two-dimensional plane to generate a cosmic muon imaging feature map.

[0089] The specific implementation steps of this step S104 are described as follows.

[0090] S1041, project the density distribution result of the closed block obtained by step S103 into a two-dimensional plane density distribution result;

[0091] In this embodiment, the two-dimensional plane is specifically a horizontal plane, and the density distribution result of each pixel on the two-dimensional plane is the density accumulation result of the longitudinal voxel.

[0092] S1042, based on the minimum density value and the maximum density value in the two-dimensional plane density distribution result, normalize the density value of each pixel on the two-dimensional plane to the interval [0, 1].

[0093] S1043, substitute the normalized density value of each pixel on the two-dimensional plane into the color mapping formula to obtain the color value of each pixel on the two-dimensional plane; wherein the color mapping formula is:

[0094] C=round(a×255) + 1,

[0095] In the formula, C is the color index of the pixel, a is the normalized density value of the pixel, and round(·) represents taking the positive integer of the calculation result.

[0096] S1045, based on the two-dimensional plane coordinate system of the closed block 1 on the horizontal plane, generate a two-dimensional cosmic muon imaging feature map.

[0097] In this embodiment, this step can be implemented by MATLAB software, and the specific operations are: 1) calling the jet function to generate a 256×3 matrix to represent the jet color mapping; wherein each row of the matrix represents a color, the first column is the red component, the second column is the green component, and the third column is the blue component; 2) creating a density matrix, the density values ​​in which are the normalized density results of each pixel on the two-dimensional plane; 3) setting the color mapping formula: C=round(a×255) + 1, a is substituted into the normalized density value of the pixel, and the integer interval is converted to [1,256] to correspond to the 256 rows of the color mapping table, each row representing a specific color; 4) calling the ind2rgb function to convert the indexed image into an RGB image.

[0098] After the above steps, the density distribution results of each pixel on the two-dimensional plane are mapped to the color distribution results; specifically, if the pixel color value is a red hue, it means that the pixel corresponds to a low-density interval; if the pixel color value is a green hue, it means that the pixel corresponds to a medium-density interval; if the pixel color value is a blue hue, it means that the pixel corresponds to a high-density interval.

[0099] In this embodiment, the two-dimensional plane coordinate system of the closed block 1 takes the center point of the block as the origin, its length direction is the x-axis direction, and its width direction is the y-axis direction. Figure 4 Shown is a cosmic muon imaging feature map generated in the scenario of this embodiment.

[0100] S2. Build and train a core defect detection network to identify defect areas on the cosmic muon imaging feature map generated in step S1 and output a defect probability map to determine the defect areas of the imaged object.

[0101] The specific implementation steps of step S2 are described in detail as follows.

[0102] S201. Build a core defect detection network.

[0103] See also Figure 5 ,In this step S201, the core defect detection network is composed of an encoder and a decoder;

[0104] The encoder consists of the first, second, third, and fourth convolutional layers connected in sequence. Each convolutional layer uses a 3×3 convolutional layer with a stride of 2. The encoder extracts features layer by layer through four convolutional layers, and gradually increases the number of image channels from 64 to 512.

[0105] The decoder is composed of a first deconvolution layer, a second deconvolution layer, a third deconvolution layer and a fourth deconvolution layer; each deconvolution layer adopts a 4x4 convolution layer with a step of 2; wherein the output end of the first convolution layer is connected with the output end of the third deconvolution layer and the input end of the first splicing module, and the output end of the first splicing module is connected with the input end of the fourth deconvolution layer; the output end of the second convolution layer and the output end of the second deconvolution layer are connected with the input end of the second splicing module, and the output end of the second splicing module is connected with the input end of the third deconvolution layer; the output end of the third convolution layer and the output end of the first deconvolution layer are connected with the input end of the third splicing module, and the output end of the third splicing module is connected with the input end of the second deconvolution layer; the decoder restores details layer by layer and finally outputs a defect probability map (512x512x1) with the same size as the input image.

[0106] S202, construct a cosmic muon image data set for core defect detection network training.

[0107] The specific implementation steps of this step S202 are:

[0108] 1) The same method as step S1 is used to obtain a plurality of cosmic muon imaging feature maps without defects on the imaging object in the closed block, and a plurality of cosmic muon imaging feature maps with known core defects of different shapes and different types on the imaging object; wherein the core defects include but are not limited to cracks, holes, and corrosion; the cosmic muon imaging feature maps with defects on the imaging object are not less than 100;

[0109] 2) The cosmic muon imaging feature maps obtained in step 1) are uniformly adjusted to images with the same resolution;

[0110] 3) According to the actual defect position on the imaging object, pixel-level manual labeling is performed on the cosmic muon imaging feature maps using the LabelMe tool, and the defects are accurately labeled and assigned with defect category labels; a plurality of cosmic muon imaging feature maps without defects on the imaging object are directly assigned with a label of no defect category; then, the plurality of cosmic muon imaging feature maps in step 1) are converted into a plurality of cosmic muon imaging feature marked images to form a cosmic muon image data set.

[0111] In this embodiment, the core defect detection network of step S201 is constructed based on the U-Net model of Python software; in step S202, the images with the same resolution specifically refer to images with a resolution of 512x512; further, the image input end of the core defect detection network is set to input 512x512x3 images, and finally the image output end is set to output 512x512x1 defect probability map.

[0112] S203 , using the cosmic muon image dataset constructed in step S202 to train the core defect detection network constructed in step S201 .

[0113] In step S203, the cosmic muon image dataset is split into a training set, a validation set, and a test set in a ratio of 8:1:1. During the training process, the Adam optimizer is used and the initial learning rate is set to 1e -4 , the decay strategy is dynamic decay; the number of iterations is set to 50 rounds; the loss function L Combining Dice Loss and BCE Loss to alleviate the problem of class imbalance, the expression is:

[0114] ,

[0115] Where, L Dice is the Dice loss function, L BCE is the BCE loss function, λ is the weight coefficient, and its value is .

[0116] The expression of Dice loss function is:

[0117] ,

[0118] Where, is the overlapping area (dot product) between the predicted defect area and the actual defect area, is the total area of ​​the actual defect area, is the total area of ​​the predicted defect region, is the smoothing constant (in this embodiment, 10 -6 ), to prevent the denominator from being zero.

[0119] The expression of the BCE loss function is:

[0120] ,

[0121] Where, N is the total number of pixels in the image; y i For the i The true label of each pixel, 0 = background, 1 = defect; p i The model predicts i The probability that a pixel is defective, p i ∈[0,1].

[0122] S204, the cosmogenic muon imaging feature map obtained in step S1 is also adjusted to an image with a resolution of 512x512, and then input into the trained core defect detection network, and a 512x512x1 defect probability map is output.

[0123] In this embodiment, the cosmogenic muon imaging feature map as shown in Figure 4 is input into the core defect detection network trained in step S2, and a black and white defect probability map is obtained. Based on the defect probability map, the pixels with defects on the imaging object can be identified. As shown in Figure 6 , a schematic diagram of marking the defect position on the imaging object on the cosmogenic muon imaging feature map with a red rectangular frame is shown. The defect position is the position for ultrasonic imaging in step S3. Figure 4

[0124] S3, according to the recognition result of step S2, ultrasonic imaging is performed on the defect area of the imaging object.

[0125] In order to accurately image and identify the defect area of the imaging object based on the cosmogenic muon imaging feature map, based on the recognition result of step S3, ultrasonic imaging is continued on the defect area of the imaging object.

[0126] The specific implementation steps of this step S3 are described as follows.

[0127] S301, referring to Figure 7 , an ultrasonic imaging detector 4 is arranged on the back of the closed block 1, and the probe of the ultrasonic imaging detector 4 is directed towards the defect area of the imaging object 2. A suitable probe is selected according to the thickness of the imaging object in the detection direction. In the probe selection, the imaging object with a thickness ≤20mm in the detection direction is defined as a thin object, and a high-frequency probe is selected accordingly. The imaging object with a thickness >20mm in the detection direction is defined as a thick object, and a low-frequency probe is selected accordingly.

[0128] In this embodiment, the ultrasonic imaging detector 4 is a SyncScan 32P ultrasonic imaging detector of Shanchao, which is arranged on the back of the closed block 1. Based on the thickness of the imaging object 2 in the detection direction being 50mm, a low-frequency probe is selected.

[0129] S302, the ultrasonic parameters suitable for ultrasonic imaging of the imaging object in the closed block are set, including: gain, detection frequency, depth range, scanning speed and gate. Specifically,

[0130] ​1) gain, which is initially set to 40dB, and in subsequent ultrasonic detection, according to the signal situation in the range of ±20dB fine-tuning; this parameter is used to adjust the amplification multiple of the received signal, in actual detection, the value is too large or too small, which cannot effectively obtain the detection signal of the object in the closed space; in this embodiment, the gain is set to 40dB, and there is no adjustment in the subsequent;

[0131] 2) detection frequency, which is based on the definition of the size of the object to be detected in the detection direction, the detection frequency of the thin object is set to 5MHz~10MHz, and the detection frequency of the thick object is set to 2MHz~4MHz; this parameter determines the wavelength and penetration ability of the ultrasonic wave; in this embodiment, based on the thickness of the imaging object 2 in the detection direction is 50mm, the detection frequency is set to 5MHz;

[0132] 3) depth range, which is set according to the size of the object to be detected in the detection direction, specifically depth range = thickness + 10mm~15mm; in this embodiment, based on the thickness of the imaging object 2 in the detection direction is 50mm, the depth range is set to 60mm to completely display the internal situation;

[0133] 4) scanning speed, which is set to 10mm / s~50mm / s; this parameter affects the imaging speed and quality; in this embodiment, the ultrasonic scanning speed is set to 40mm / s;

[0134] 5) gate, which is set according to the distance between the ultrasonic imaging detector and the closed space, so that the starting position of the detection is in the closed space, and the detection width is set according to the range of the defect area, so that the detection width covers the defect area width, and the height is adjusted between 2mm~10mm according to the signal amplitude; this parameter is used to analyze the signal of a specific area; in this embodiment, the starting position in the gate setting is set in the closed block and is 5mm away from the back surface of the imaging object, the width is set to 20mm, which exceeds the defect area width identified in step S2, and the height is set to 6mm.

[0135] S303, acquiring an ultrasonic image of the defect area of the imaging object by using the ultrasonic imaging detector.

[0136] In this step, when the ultrasonic imaging detector emits ultrasonic waves, the reflected waves are generated due to the difference in acoustic impedance when the ultrasonic waves encounter defects or material interfaces. The intensity and time delay signals of the reflected waves are detected, and by processing and analyzing the signals, an ultrasonic high-precision image identifying the defective state can be output. However, the Doppler ultrasound can currently only encode the internal motion information of the target object (such as fluid flow, tissue vibration, etc.) based on the Doppler frequency shift principle, and through a specific algorithm, different frequency shifts are corresponded to different color channels to form a color image. The application scenario of the embodiment is the static feature detection of the surface and internal defects of the target object, and these static defects hardly cause or only cause extremely small ultrasonic frequency shifts, making it difficult for Doppler ultrasound to effectively perceive and extract feature signals related thereto; therefore, the ultrasonic imaging image is a grayscale image, and the ability to confirm the specific distribution of defects on the imaged object is limited, and since the ultrasonic image only has relative position information, it cannot be directly fused with the primordial muon imaging feature map obtained in step S1.

[0137] As shown in Figure 8 is a grayscale ultrasonic image preliminarily obtained by using an ultrasonic imaging detector. From the ultrasonic image, the specific distribution of defects on the imaged object cannot be clearly identified.

[0138] S304, based on the original ultrasonic grayscale image, ultrasonic image reconstruction is performed to obtain a color image that can be fused with the primordial muon image.

[0139] This step S304 can be directly realized by using the processing software (Tomoview software) matched with the ultrasonic imaging detector, and the specific implementation steps are described as follows.

[0140] S3041, feature extraction and quantization: in the instrument parameter setting menu, the characteristic parameter is selected as the echo amplitude, and it is normalized to the value range interval of 0~255, so as to facilitate the color mapping setting.

[0141] S3042, color mapping setting: for the echo amplitude value normalized to the range of 0~255, different thresholds are set to map to different colors; specifically, 0~30 is set as dark blue and represents large defects, 31~80 is set as light blue and represents small defects, 81~200 is set as green and represents normal areas, and 201~255 is set as red and represents strong reflection interfaces; for each color, the depth threshold is set at an interval of 10 mm, so that the ultrasonic pixels increase in depth with the increase of the quantized value, and the color saturation also increases with the increase of the depth, so as to distinguish different depth structure information.

[0142] S3043, image reconstruction setting: based on the echo amplitude normalization data obtained by step S3041 and the color mapping setting of step S3042, image reconstruction is performed; wherein the parameter setting of the image reconstruction process is: the scanning mode is selected as sector scanning, the scanning angle is set to 60°, the scanning depth is consistent with the thickness of the imaging object in the detection direction, the scanning step is set to 0.5mm, and the reconstruction algorithm is selected as the image reconstruction algorithm based on the coherence factor; further, through this step, a high-resolution color image is reconstructed, which makes the defects more easily identifiable.

[0143] S3044, image post-processing setting: running Sobel edge enhancement algorithm in the ultrasound software to process the image reconstructed by step S3043; wherein in the Sobel edge enhancement algorithm module, the intensity parameter is set to 3 to highlight the edges and avoid image distortion; Gaussian smoothing algorithm is selected for denoising, and the smoothing degree parameter is set to 2 to remove noise while retaining image details; histogram equalization algorithm is selected, and the contrast parameter is set to 1.5 times to enhance the contrast and make the display of different regions in the image more clear.

[0144] After the above step S304, the initially obtained gray-scale ultrasound image is reconstructed into a color image that can display defects of different degrees, as shown in Figure 9 From Figure 9 it can be seen that there is an obvious E-shaped irregular crack on the imaging object in the ultrasonic detection area, which is consistent with the known defects on the imaging object B part when the scene is set in this embodiment.

[0145] S4, the primordial muon imaging feature map obtained by step S1 and the ultrasound image obtained by step S3 are fused to realize detailed local imaging of the defect area based on the primordial muon imaging feature map, and a fused image is obtained, which clearly shows the position and shape of the defect of the imaging object in the closed block.

[0146] As described in step S3 above, the initial gray-scale ultrasound image is reconstructed into a color image that can display defects of different degrees, but since there is no absolute position information, it cannot be directly fused to the corresponding position of the primordial muon imaging feature map, so it is necessary to determine the accurate fusion position of the ultrasound image on the primordial muon imaging feature map.

[0147] This step S4 can be realized by OpenCV library of Python software, and the implementation steps are as follows:

[0148] S401, obtaining a non-defect area mask of the primordial muon imaging feature map, and the specific steps are as follows:

[0149] 1) Convert the probability map output by step S2 into a binary mask; specifically, define the value of a pixel with a probability ≥ 0.5 as 1 and the value of a pixel with a probability < 0.5 as 0 with a threshold of 0.5;

[0150] 2) Multiply the value of each pixel in the binary mask by 255 and convert it into uint8 type to facilitate OpenCV morphological operation;

[0151] 3) Define: the morphological closing operation uses a 3x3 rectangular kernel, and the structure element is defined as:

[0152]

[0153] The mathematical expression of the closing operation is:

[0154]

[0155] wherein, represents the dilation operation, ; represents the erosion operation, ;

[0156] 4) Perform the closing operation on the mask obtained in step 2) using the method defined in step 3), and set the number of iterations of the closing operation to 2 (i.e., perform the closing operation twice) to enhance the continuity of the defect area;

[0157] 5) Divide the value of each pixel in the mask obtained in step 4) by 255 to output a new binary mask; wherein in the new binary mask, the pixel with a value of 1 corresponds to a defect pixel;

[0158] 6) Perform a mask inversion operation on the new binary mask obtained in step 5) using the mask inversion operation module (i.e., cv2.bitwise_not) to obtain a color mask of the non-defect area of the extracted muon imaging feature map;

[0159] S402, obtain a defect area mask of the ultrasound color image, and the specific steps are as follows:

[0160] 1) Adjust the ultrasound color image to the same resolution image as the muon imaging feature map; in this embodiment, the image resolution is 512x512;

[0161] 2) Extract the defect area mask in the ultrasound color image after adjusting the resolution in step 1) using the mask extraction module (i.e., cv2.bitwise_and);

[0162] S403, image fusion, and the specific steps are as follows:

[0163] ​​Using an image fusion module (i.e., cv2.bitwise_or), the non-defective area mask obtained in step S401 and the defective area mask obtained in step S402 are fused;

[0164] The weighted mixing formula of the fusion process of the image fusion module is:

[0165] I fused = I RGB非缺陷 ×(1- α ) + I US缺陷 × α ,

[0166] Where, I RGB非缺陷 is the intensity of the non-defective area of ​​the RGB image; I US缺陷 is the intensity of the defect area in the ultrasonic image; α is the weight, and its value range is 0.4~0.8; in this embodiment, the weight parameter α is set to 0.6 in the image fusion module.

[0167] like Figure 10 The figure shows the fusion image obtained by fusing the ultrasound color images obtained in step S1 and the ultrasound color images obtained in step S4 using the fusion processing step above. Figure 10 In the fusion image, not only can the presence of an E-shaped crack in part B of the imaged object be clearly determined, but the exact location of the crack in part B can also be clearly determined. This imaging result provides effective image support for subsequent treatment of local defects in the imaged object.

[0168] In summary, the high-precision ultrasonic imaging recognition method based on cosmic muon images of the present invention not only retains the overall rough image of the target object, but also can effectively perform high-precision and lossless imaging of the defective areas of the imaging object. Moreover, the imaging process is completely radiation-free and has a good prospect for application and promotion.

Claims

1. A high-precision ultrasonic imaging recognition method based on cosmic muon images, characterized in that: Here are the steps: S1. Lay two liquid scintillator muon detectors, one on top and one on the bottom, on the closed block with the built-in imaging object to collect the true track information of each muon entering and exiting the closed block to generate the cosmic muon imaging feature map; S2. Build and train a core defect detection network to identify defect regions on the cosmic muon imaging feature map generated in step S1 and output a defect probability map to determine the defect regions of the imaged object. S3. Based on the recognition result of step S2, ultrasonic imaging is performed on the defective area of ​​the imaging object. Then, the echo amplitude in the ultrasonic imaging is normalized to the interval [0, 255] and color mapping is performed to obtain an ultrasonic color image that can be integrated with the cosmic muon imaging feature map. S4. Obtain the non-defect area mask of the cosmic muon imaging feature map and the defect area mask of the ultrasonic color image respectively. By fusing the two images, a fused image is obtained that clearly shows the defect position and shape of the imaging object in the closed block.

2. The high-precision ultrasonic imaging recognition method based on cosmic muon images according to claim 1 is characterized in that: The specific implementation steps of step S1 are as follows: S101, obtaining real muon scattering angle data of the closed block based on the real track information of each muon entering the closed block and the real track information of each muon exiting the closed block; S102, constructing a forward model to obtain theoretical muon scattering angle data of a closed block by simulating muon flux distribution; S103, calculating the optimal solution of density distribution of the closed block by using the MLSD inversion algorithm; S104. Project the optimal solution of the density distribution of the closed block onto a two-dimensional plane and assign color values ​​to generate a cosmic muon imaging feature map.

3. The high-precision ultrasonic imaging recognition method based on cosmic muon images according to claim 2 is characterized in that: Step S102 is implemented using Geant4 software; the modeling of the enclosed block and the imaging object is consistent with reality, the model of the object to be imaged is set to have a uniform theoretical density value, and the remaining space in the enclosed block is set to be filled with air. A liquid scintillator muon detector is set in the center of the upper and lower paths of the enclosed block model so that the scattering data collection area completely covers the enclosed block model, and the direction of the cosmic emission muons is set to random.

4. The high-precision ultrasonic imaging recognition method based on cosmic muon images according to claim 3 is characterized in that: The specific processing steps of step S103 are: S1031, dividing the closed block 1 into a voxel grid, and assigning an initial density value to each voxel; S1032. Define the objective function F: , where △ θ i For the i The actual scattering angle of a muon, △ θ' i For the i The theoretical scattering angle of muons, N is the total number of muons; S1033. Using the least squares method, with minimizing the objective function F as the optimization goal, the optimal density distribution result of the closed block is obtained through continuous iterative solution.

5. The high-precision ultrasonic imaging recognition method based on cosmic muon images according to claim 1 is characterized in that: The specific implementation steps of step S2 are: S201. Construct a core defect detection network. The network encoder consists of a first convolutional layer, a second convolutional layer, a third convolutional layer, and a fourth convolutional layer connected in sequence. Each convolutional layer uses a 3×3 convolutional layer with a stride of 2. The decoder consists of a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, and a fourth deconvolution layer; each deconvolution layer adopts a 4×4 convolution layer with a stride of 2; wherein the output end of the first convolution layer and the output end of the third deconvolution layer are connected to the input end of the first splicing module, and the output end of the first splicing module is connected to the input end of the fourth deconvolution layer; the output end of the second convolution layer and the output end of the second deconvolution layer are connected to the input end of the second splicing module, and the output end of the second splicing module is connected to the input end of the third deconvolution layer; the output end of the third convolution layer and the output end of the first deconvolution layer are connected to the input end of the third splicing module, and the output end of the third splicing module is connected to the input end of the second deconvolution layer; S202, constructing a cosmic muon image dataset for core defect detection network training, using the same method as step S1, respectively obtaining several cosmic muon imaging feature maps of defect-free objects in a closed block, and several cosmic muon imaging feature maps of objects with known core defects of different shapes and types, and uniformly adjusting the resolution, performing pixel-level defect annotation on the cosmic muon imaging feature maps, and assigning defect category labels to form a dataset; S203, training a core defect detection network using a cosmic muon image dataset; S204: After adjusting the cosmic muon imaging feature map obtained in step S1 to an image of the same resolution, input it into the trained core defect detection network and output a defect probability map.

6. The high-precision ultrasonic imaging recognition method based on cosmic muon images according to claim 5 is characterized in that: During the training process in step S203, the Adam optimizer is used and the initial learning rate is set to 1e -4 , the decay strategy is dynamic decay; the number of iterations is set to 50 rounds; the loss function L Combining Dice Loss and BCE Loss to alleviate the problem of class imbalance, the expression is: ， Where, L Dice is the Dice loss function, L BCE is the BCE loss function, λ is the weight coefficient, and its value is .

7. The high-precision ultrasonic imaging recognition method based on cosmic muon images according to claim 1 is characterized in that: In step S3, ultrasonic imaging is achieved by an ultrasonic imaging detector, which is set at the front or back of the closed block; wherein, Select the probe based on the thickness of the imaging object in the detection direction. If the thickness of the imaging object in the detection direction is ≤20mm, it is considered a thin object and a high-frequency probe is used. If the thickness of the imaging object in the detection direction is greater than 20mm, it is considered a thick object and a low-frequency probe is used. Set ultrasound parameters suitable for ultrasound imaging of objects within a closed area, including: 1) Gain, which is initially set to 40dB and fine-tuned within the range of ±20dB during subsequent ultrasonic testing; 2) Detection frequency: the detection frequency for thin objects is set to 5MHz~10MHz, and the detection frequency for thick objects is set to 2MHz~4MHz; 3) Depth range, which is set to increase the thickness of the imaged object by 10 mm to 15 mm in the detection direction; 4) Scanning speed, which is set to 10mm / s~50mm / s; 5) Gate, which is set so that the starting position of detection is in a closed space, the detection width is the width of the covered defect area, and the detection height is 2mm~10mm.

8. The high-precision ultrasonic imaging recognition method based on cosmic muon images according to claim 1 is characterized in that: The image reconstruction steps of step S3 are: 1) Select the echo amplitude of ultrasound imaging as the characteristic parameter and normalize it to a numerical range of 0 to 255; 2) For the normalized value range, 0-30 is set as dark blue to represent large defects, 31-80 is set as light blue to represent small defects, 81-200 is set as green to represent normal areas, and 201-255 is set as red to represent strong reflective interfaces. For each color, a depth threshold is set at 10mm intervals so that the depth of the ultrasound pixel increases as the quantization value increases. 3) Based on the normalized data from step 1) and the color mapping settings from step 2), an image reconstruction algorithm based on a coherence factor is used to complete the image reconstruction; 4) Use the Sobel edge enhancement algorithm to post-process the color image reconstructed in step 3).

9. The high-precision ultrasonic imaging recognition method based on cosmic muon images according to claim 1, characterized in that: In step S4, the specific method for obtaining the non-defect area mask of the cosmic muon imaging feature map is as follows: 1) Convert the defect probability map output by step S2 into a binary mask; 2) Multiply the value of each pixel in the binary mask by 255 and convert it to uint8 type; 3) Define the morphological closing operation using a 3×3 rectangular kernel and the closing operation expression; 4) Using the method defined in step 3), perform a closing operation on the mask obtained in step 2), and set the number of closing operation iterations to 2; 5) Divide the value of each pixel in the mask obtained in step 4) by 255 and output a new binary mask; 6) Perform a mask inversion operation on the new binary mask obtained in step 5) to obtain a color mask of the non-defect area of ​​the cosmic muon imaging feature map.

10. The high-precision ultrasonic imaging recognition method based on cosmic muon images according to claim 1, characterized in that: In step S4, the specific steps of obtaining the defect area mask of the ultrasonic color image and fusing it with the non-defect area mask of the cosmic muon imaging feature map are as follows: 1) Adjust the ultrasound color image to an image with the same resolution as the cosmic muon imaging feature map; 2) Using the mask extraction module, extract the defect area mask from the ultrasound color image; 3) Fuse the non-defective area mask obtained in step 1) and the defective area mask obtained in step 2); the weighted mixing formula of the fusion process is: I fused = I RGB非缺陷 ×(1- α ) + I US缺陷 × α ， Where, I RGB非缺陷 is the intensity of the non-defective area of ​​the RGB image; I US缺陷 is the intensity of the defect area in the ultrasound image; α is the weight, and its value range is 0.4~0.8.

Images