Ultrasonic microscopic-based metal surface micro-defect identification method and system

Abstract

The application provides a metal surface micro-defect recognition method and system based on ultrasonic microscopy. The method is applied to a recognition system and comprises the following steps: collecting an image of a metal surface to be recognized by using an ultrasonic signal with a target frequency; extracting at least one feature to be recognized from the image, wherein any feature to be recognized contains a defect feature of the metal surface to be recognized; when the at least one feature to be recognized contains two or more features to be recognized, the defect features contained in any two features to be recognized are different in size; recognizing the at least one feature to be recognized respectively to obtain at least one defect recognition sub-result; and combining the at least one defect recognition sub-result to obtain a defect recognition result of the metal surface to be recognized. By using the technical solution, the detection rate of micron-level defects can be improved, and the accuracy of defect recognition can be improved by using a deep learning model to recognize defects.

Description

Technical Field

[0001] The present application belongs to the field of non-destructive testing technology, and specifically relates to a method and system for identifying metal surface micro-defects based on ultrasonic microscopy. Background Art

[0002] Surface defect identification of metal materials is a necessary process for applying metal materials to industrial production. With the development of industrialization, the requirement for the recognition accuracy of surface defects of metal materials can reach the micron level.

[0003] Currently, the primary method for detecting defects on metal surfaces involves scanning the metal with an instrument to obtain a microscopically magnified image of the surface, and then manually identifying defects in the image. However, in actual implementation, due to instrument performance limitations, the resulting microscopically magnified image lacks sufficient visibility of micron-level defects. Furthermore, manual identification relies heavily on experience, resulting in low detection accuracy and a high rate of missed detections. Summary of the Invention

[0004] This application proposes a method and system for identifying metal surface micro-defects based on ultrasonic microscopy, which can not only improve the detection rate of micron-level defects, but also use a deep learning model to identify defects, thereby improving the recognition accuracy.

[0005] The first embodiment of the present application proposes a method for identifying metal surface micro-defects based on ultrasonic microscopy, which is applied to an identification system and includes:

[0006] Acquiring an image of the metal surface to be identified using an ultrasonic signal of a target frequency, wherein the target frequency is greater than a preset frequency threshold;

[0007] extracting at least one feature to be identified from the image, any feature to be identified including defect features of the metal surface to be identified; when the at least one feature to be identified includes two or more features to be identified, the scales of the defect features included in any two features to be identified are different;

[0008] Respectively identifying the at least one feature to be identified to obtain at least one defect identification sub-result, wherein the at least one feature to be identified corresponds to the at least one defect identification sub-result in a one-to-one manner;

[0009] The at least one defect recognition sub-result is combined to obtain a defect recognition result of the metal surface to be recognized, wherein the defect recognition result includes at least one of a defect type, a defect position, and a defect size.

[0010] In some embodiments of the present application, the target frequency is 300 MHz, and the step of acquiring an image of the metal surface to be identified using an ultrasonic signal at the target frequency includes:

[0011] emitting an ultrasonic signal at a frequency of 300 MHz toward the metal surface to be identified according to a scanning trajectory matching the metal surface to be identified;

[0012] For any scanning position of the metal surface to be identified, receiving a reflected wave signal of the scanning position to the ultrasonic signal;

[0013] According to the preset correspondence between the signal intensity and the grayscale value of the reflected wave signal, the grayscale value of the imaging pixel corresponding to the scanning position is determined to obtain the image of the metal surface to be identified.

[0014] In some embodiments of the present application, before acquiring an image of the metal surface to be identified using an ultrasonic signal of a target frequency, the method further includes:

[0015] In response to the creation instruction, calling a pre-deployed simulation model to create a metal sample model, and determining sample parameters of the metal sample model, wherein the sample parameters include positions and sizes of surface defects in the metal sample model;

[0016] receiving simulation parameters, the simulation parameters including the frequency of the ultrasonic signal, the ultrasonic signal transmission boundary of the metal sample model, and grid parameters, the grid parameters being used to divide the metal sample model into a plurality of grid units;

[0017] Calculating the maximum time step corresponding to the simulation model according to the grid parameters;

[0018] Calling the simulation model to perform ultrasonic microscopy simulation on the metal sample model based on the simulation parameters to determine a corresponding relationship between the signal intensity of the reflected wave signal and the defect depth;

[0019] According to the corresponding relationship between the signal intensity of the reflected wave signal and the defect depth, the corresponding relationship between the signal intensity of the reflected wave signal and the grayscale value is determined.

[0020] In some embodiments of the present application, extracting at least one feature to be identified from the image includes:

[0021] extracting high-level features and low-level features of the image respectively, wherein the high-level features include high-level features of the defects on the metal surface to be identified, and the low-level features include low-level features of the defects on the metal surface to be identified;

[0022] splicing the high-level features and the low-level features to obtain image features of the metal surface to be identified;

[0023] The at least one feature to be recognized is extracted from the image features.

[0024] In some embodiments of the present application, respectively extracting high-level features and low-level features of the image includes:

[0025] Invoking a first feature extraction network to extract high-level features of the image, the first feature extraction network comprising a first number of cascaded feature extraction modules;

[0026] A second feature extraction network is called to extract low-level features of the image, where the second feature extraction network includes a second number of feature extraction modules, and the first number is greater than the second number.

[0027] In some embodiments of the present application, when the at least one feature to be identified includes two or more features to be identified, extracting the at least one feature to be identified from the image features includes:

[0028] Upsampling the image feature N times to obtain N enhanced features of the image feature, where N is an integer greater than or equal to 2;

[0029] Performing N downsampling on the enhanced features obtained by the N-th upsampling to obtain the N reduced-dimensional features respectively;

[0030] For any dimensionality reduction feature, the dimensionality reduction feature and the enhanced feature with the same scale as the dimensionality reduction feature are concatenated to obtain the feature to be identified.

[0031] In some embodiments of the present application, the step of respectively identifying the at least one feature to be identified to obtain at least one defect identification sub-result includes:

[0032] For any feature to be identified, a recognition module that matches the scale of the feature to be identified is called to identify defect features in the feature to be identified, so as to obtain a defect identification sub-result corresponding to the feature to be identified.

[0033] In some embodiments of the present application, the at least one feature to be identified includes a feature to be identified at a first scale, a feature to be identified at a second scale, and a feature to be identified at a third scale, wherein the first scale is larger than the second scale, and the second scale is larger than the third scale;

[0034] The step of calling a recognition module that matches the scale of any feature to be recognized to recognize the defect feature in the feature to be recognized includes:

[0035] Invoking a first recognition module to recognize the feature to be recognized at the first scale, and obtaining a defect recognition sub-result of the first scale of the metal surface to be recognized;

[0036] Invoking a second recognition module to recognize the feature to be recognized at the second scale, and obtaining a defect recognition sub-result of the second scale of the metal surface to be recognized;

[0037] The third recognition module is called to recognize the feature to be recognized at the third scale, and a defect recognition sub-result of the third scale of the metal surface to be recognized is obtained.

[0038] In some embodiments of the present application, before extracting at least one feature to be identified from the image, the method further includes:

[0039] Generating a training sample set, wherein each training sample includes a label, wherein the label is used to mark whether the training sample contains a surface defect, and if the training sample contains a surface defect, the location, size, and category of the surface defect;

[0040] The training sample set is used to train a deep learning model to obtain a defect recognition model, and the defect recognition model is used to input the image and output the defect recognition result.

[0041] In some embodiments of the present application, generating a training sample set includes:

[0042] Scanning a plurality of metal samples with an ultrasonic signal of the target frequency to obtain a plurality of initial training sample images;

[0043] For any initial training sample image with surface defects, in response to a labeling instruction, a label is added to each surface defect of the initial training sample image, where each label is used to mark the position, size, and category of the corresponding surface defect;

[0044] Each M labeled initial training sample images are spliced ​​into a training sample to obtain the training sample set, where M is an integer greater than or equal to 2.

[0045] In some embodiments of the present application, the training of a deep learning model using the training sample set to obtain a defect recognition model includes:

[0046] Calling the deep learning model to predict the input training sample, and outputting the predicted position, predicted size and predicted category of the surface defect of the training sample;

[0047] Calculating a loss function, the loss function being a weighted sum of positioning loss, classification loss, and confidence loss; the positioning loss representing the loss between the predicted position and the position labeled with the training sample, the classification loss representing the loss between the predicted category and the category labeled with the training sample, and the confidence loss representing the cross entropy loss between the predicted category and the category labeled with the training sample;

[0048] If the loss function satisfies the convergence condition, determining the deep learning model as the defect recognition model;

[0049] If the loss function does not meet the convergence condition, a new training sample is determined from the training sample set, and the step of calling the deep learning model to predict the input training sample is performed again for the new training sample.

[0050] The embodiment of the second aspect of the present application provides a metal surface micro-defect recognition system based on ultrasonic microscopy, the recognition system includes a computer device and an ultrasonic microscopy detection device, and the defect recognition model is deployed in the computer device, wherein:

[0051] The ultrasonic microscopy detection device is configured to receive a trigger instruction from the computer device, capture an image of the metal surface to be identified using an ultrasonic signal at a target frequency greater than a preset frequency threshold, and transmit the image to the computer device;

[0052] The computer device is used to call the defect recognition model to extract at least one feature to be recognized from the image, any feature to be recognized includes the defect feature of the metal surface to be recognized; when the at least one feature to be recognized includes two or more features to be recognized, the scales of the defect features included in any two features to be recognized are different; the at least one feature to be recognized is recognized separately to obtain at least one defect recognition sub-result, and the at least one feature to be recognized corresponds one-to-one to the at least one defect recognition sub-result; and the at least one defect recognition sub-result is combined to obtain a defect recognition result of the metal surface to be recognized, and the defect recognition result includes at least one of the defect type, defect location and defect size.

[0053] In some embodiments of the present application, the ultrasonic microscopy detection device includes a pulse transceiver, a scanner, an ultrasonic transducer and a data collector, wherein:

[0054] The scanner is configured to respond to the trigger instruction and move the ultrasonic transducer along a scanning trajectory that matches the metal surface to be identified;

[0055] The pulse transceiver is used to send an excitation signal to the ultrasonic transducer in response to the trigger instruction;

[0056] The ultrasonic transducer is configured to transmit an ultrasonic signal at a frequency of 300 MHz to the metal surface to be identified in response to the excitation signal; and is further configured to receive a reflected wave signal of the ultrasonic signal at any scanning position on the metal surface to be identified;

[0057] The data collector is used to determine the grayscale value of the imaging pixel corresponding to the scanning position according to the preset corresponding relationship between the signal intensity and the grayscale value of the reflected wave signal, so as to obtain the image of the metal surface to be identified.

[0058] The technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0059] The recognition system of the embodiment of the present application uses an ultrasonic signal with a target frequency greater than a preset frequency threshold to capture images of the metal surface to be identified. Since the higher the frequency of the ultrasonic signal, the stronger the ability to detect micron-level defects, based on this, the image captured using a higher-frequency ultrasonic signal can, to a certain extent, improve the detection rate of smaller micron-level defects, so that the captured image can more clearly show the characteristics of smaller defects. Furthermore, the recognition system extracts the features of the image and identifies the defects of the metal surface to be identified based on the features of the image. Not only does it not require manual identification, but it also uses a deep learning model to identify defects, which can improve the accuracy of identification and reduce the missed detection rate.

[0060] Additional aspects and advantages of the present application will be given in part in the description below and in part will become apparent from the description below or learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the optional embodiments below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present application. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0062] Figure 1A FIG1 shows an exemplary architecture diagram 1 of a metal surface micro-defect recognition system based on ultrasonic microscopy provided in one embodiment of the present application;

[0063] Figure 1B The schematic diagram shows an exemplary architecture of a metal surface micro-defect recognition system based on ultrasonic microscopy provided in an embodiment of the present application. Figure 2 ;

[0064] Figure 1C A schematic diagram illustrating an exemplary scenario of an ultrasonic signal propagation scenario provided by an embodiment of the present application is shown;

[0065] Figure 2 A flowchart of a method for identifying metal surface micro-defects based on ultrasonic microscopy provided in one embodiment of the present application is shown;

[0066] Figure 3A A schematic diagram illustrating an exemplary structure of a defect recognition model provided in an embodiment of the present application is shown;

[0067] Figure 3B Shown Figure 3A An exemplary structural diagram of a central backbone network;

[0068] Figure 3C Shown Figure 3A Schematic diagram of the exemplary structure of the mid-neck network and prediction network;

[0069] Figure 4A A diagram showing a comparison of three defect recognition results provided by an embodiment of the present application; and a schematic diagram of a storage medium;

[0070] Figure 4B A comparison diagram of ultrasonic microscopy imaging and electron microscopy imaging in the present invention is shown. DETAILED DESCRIPTION

[0071] The following describes exemplary embodiments of the present application in more detail with reference to the accompanying drawings. Although exemplary embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. Instead, these embodiments are provided to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0072] It should be noted that, unless otherwise specified, the technical or scientific terms used in this application should have the common meanings understood by those skilled in the art to which this application belongs.

[0073] First, the technical scenarios involved in the embodiments of the present application are described.

[0074] The embodiments of the present application relate to non-destructive testing scenarios for metal surfaces. Non-destructive testing, also known as defect identification, refers to the process of locating, classifying and evaluating surface defects of metal materials, metal products, etc. using various detection technologies and methods. Defects on the metal surface may include discontinuities, anomalies or flaws on the metal surface, such as cracks, holes, inclusions, scratches, protrusions, etc. Defect identification on the metal surface plays a key role in quality control in industrial production. The main methods for defect identification on the metal surface include using an instrument to scan the metal to obtain a microscopic magnified image of the metal surface, and then identifying defects on the metal surface based on the microscopic magnified image.

[0075] Conventional methods for scanning and obtaining microscopic magnified images of metal surfaces include X-ray testing, magnetic particle testing, and eddy current testing. However, with the gradual advancement of industrialization, the detection accuracy of metal surface defects has reached the micron level, but these conventional testing methods have limited accuracy for micron-level defects. Ultrasonic microscopy, while relatively accurate, is still unable to detect smaller micron-level defects due to the current frequency of ultrasonic signals generally ranging from 15 MHz to 200 MHz. Furthermore, the current method of manually identifying metal surface defects through microscopic magnified images relies heavily on manual experience, resulting in low detection accuracy and a high rate of missed detections. Consequently, conventional metal surface defect identification methods suffer from low accuracy and high miss rates in both the detection and identification stages, resulting in the lack of efficient and high-precision identification methods for micron-level defects on metal surfaces.

[0076] In light of this, the technical solution provided in the embodiments of this application utilizes higher-frequency ultrasonic signals to capture images of the metal surface to be identified, thereby improving the detection rate of smaller micron-level defects, allowing the captured images to more clearly reveal the characteristics of smaller defects. Furthermore, the embodiments of this application utilize a deep learning model to identify defects on the metal surface to be identified, thereby improving recognition accuracy and reducing the rate of missed detections.

[0077] The following specific embodiments describe in detail the technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0078] See also Figure 1A , Figure 1A A metal surface micro-defect recognition system based on ultrasonic microscopy (hereinafter referred to as the recognition system) is shown, and the recognition system includes a computer device 10 and an ultrasonic microscopy detection device 20. The computer device 10 and the ultrasonic microscopy detection device 20 are communicatively connected.

[0079] The computer device 10 can be configured as an industrial computer. The computer device 10 may include functional components such as a processor, memory, and a communication module. The communication module can be used to provide communication between the computer device 10 and the ultrasonic microscopy device 20, enabling signal transmission and reception with the ultrasonic microscopy device 20. The memory can be used to store program instructions, data, and deep learning models that support the implementation of the technical solutions of this application. Program instructions include, for example, instructions for triggering the ultrasonic microscopy device 20 to transmit ultrasonic signals and instructions for invoking a defect recognition model. Data may include, for example, model parameters of the deep learning model and the correspondence between the signal intensity and grayscale value of the reflected wave signal. The deep learning model may be, for example, a pre-trained model for defect recognition (referred to as a defect recognition model), which may include multiple convolution modules. The processor can transmit a trigger instruction to the ultrasonic microscopy device 20 to trigger the ultrasonic microscopy device 20 to transmit ultrasonic signals to the metal surface to capture an image of the metal surface. After capturing the image of the metal surface, the processor may also invoke the defect recognition model to identify defects on the metal surface. The computer device 10 can identify micron-level defects on the metal surface to be inspected.

[0080] The ultrasonic microscopy device 20 may be configured to, upon receiving a trigger instruction from the computer device, capture an image of the metal surface to be identified using an ultrasonic signal at a target frequency, and transmit the image to the computer device. In some embodiments, the target frequency is greater than a preset frequency threshold, which may be, for example, 200 MHz. For example, the target frequency is 300 MHz.

[0081] In some embodiments, such as Figure 1B In the exemplary identification system shown, the ultrasonic microscopy detection device 20 may include: a pulse transceiver 201, a scanner 202, an ultrasonic transducer 203 and a data collector 204, wherein the scanner 202, the ultrasonic transducer 203 and the data collector 204 are all communicatively connected to the pulse transceiver 201, and the pulse transceiver 201 is communicatively connected to the computer device 10.

[0082] In actual implementation, the pulse transceiver 201 and the scanner 202 can both receive a trigger instruction from the computer device 10. In response to the trigger instruction, the pulse transceiver 201 can send an excitation signal to the ultrasonic transducer 203 to stimulate the ultrasonic transducer 203 to generate a pulsed ultrasonic wave (i.e., an ultrasonic signal). In response to the trigger instruction, the scanner 202 can move the ultrasonic transducer 203 along a scanning trajectory that matches the metal surface and scan the metal surface. For example, the scanning trajectory of the scanner 202 can be implemented as: moving from one side of the metal to the other side along the x-axis in the plane coordinate system, and determining the starting point of a new side in the x-axis direction according to the y-axis, thereby moving the ultrasonic transducer 203 to scan each position for detection, and enabling the ultrasonic transducer 203 to obtain a reflected wave signal at the scanning position.

[0083] The ultrasonic transducer 203 is configured to respond to the excitation signal, for example, to transmit an ultrasonic signal to the metal surface at a frequency of 300 MHz, and for any scanning position on the metal surface, receive a reflected wave signal of the scanning position to the ultrasonic signal. Exemplarily, a piezoelectric chip can be deployed in the ultrasonic transducer 203. The piezoelectric chip can generate a high-frequency ultrasonic wave of 300 MHz in response to the excitation signal. The ultrasonic wave can be focused by the acoustic lens and vertically incident on the metal surface. After that, a part of the ultrasonic wave is reflected at the discontinuity of the acoustic characteristics of the metal surface. The reflected wave can be received by the piezoelectric chip and transmitted to the pulse transceiver 201 to be converted into an electrical signal via the pulse transceiver 201. The other part of the ultrasonic wave is transmitted into the interior of the metal material. When there are no defects on the metal surface, there are no defects in the transmission path of the ultrasonic wave, and the ultrasonic wave can only generate a reflected wave, such as Figure 1C When there are defects on the metal surface, there are defects in the transmission path of the ultrasonic wave, which causes the ultrasonic wave to reflect or scatter echoes, as shown in (a). Figure 1C Based on this, by receiving and analyzing the reflected signal from the metal surface, the defects of the metal material can be detected.

[0084] In other embodiments, the pulse transceiver 201 may also be configured to limit and amplify the electrical signal corresponding to the reflected signal, and then transmit the amplified electrical signal to the data collector 204 .

[0085] The data collector 204 is configured to convert the electrical signal into a digital signal, and determine the grayscale value of the imaging pixel corresponding to the scanning position based on the preset correspondence between the signal intensity and grayscale value of the reflected wave signal, so that the computer device can obtain the image of the metal surface to be identified.

[0086] The following combination Figures 1A to 1CThe implementation scenario described here illustrates the metal surface micro-defect identification method based on ultrasonic microscopy of the present application.

[0087] See also Figure 2 , Figure 2 A method flow chart of a metal surface micro-defect identification method based on ultrasonic microscopy provided in an embodiment of the present application is shown. Figure 1A or Figure 1B The identification system shown. The method specifically includes the following steps:

[0088] Step S101 : collecting an image of a metal surface to be identified using an ultrasonic signal of a target frequency.

[0089] As described in the aforementioned embodiment, the target frequency is greater than the preset frequency threshold.

[0090] Taking the target frequency as 300MHz as an example, combined with Figure 1A and Figure 1B In the identification system shown, the computer device can trigger the ultrasonic microscopy detection device to transmit an ultrasonic signal to the metal surface to be identified at a frequency of 300 MHz according to a scanning trajectory that matches the metal surface to be identified, and receive a reflected wave signal of the ultrasonic signal at any scanning position of the metal surface to be identified. Furthermore, the computer device can determine the grayscale value of the imaging pixel corresponding to the scanning position based on a preset correspondence between the signal intensity and grayscale value of the reflected wave signal to obtain an image of the metal surface to be identified.

[0091] It should be noted that the principle of ultrasonic microscopic defect detection is to use the propagation characteristics of ultrasonic waves in metal materials to detect defects in metal materials. Generally, the higher the frequency f of the ultrasonic wave and the shorter the wavelength λ, the higher the resolution of the defect. Resolution refers to the minimum distance between two adjacent defects that the ultrasonic microscopic detection device can distinguish. The frequency f is related to the propagation speed C of the ultrasonic wave. When the propagation speed C is constant, the wavelength is inversely proportional to the frequency. That is, the relationship between the frequency f, wavelength λ and propagation speed C of the ultrasonic wave can satisfy: Based on this, although ultrasonic waves with a frequency range of 15 MHz to 200 MHz can detect some micron-level defects, the resolution is insufficient for smaller micron-level defects (such as a few microns to tens of microns).

[0092] By adopting this implementation method, defect detection at an ultrasonic frequency of 300 MHz can obtain a higher resolution, thereby making it possible to detect smaller micron-level defects more powerful, thereby facilitating improved detection accuracy.

[0093] For example, combined Figure 1B, taking the ultrasonic transducer as a water-immersed point-focused ultrasonic transducer as an example, using distilled water as a coupling agent, and the scanning step length is 1 micron ( ). A focused ultrasonic transducer is a flat probe with an acoustic lens installed in front of it. The acoustic lens is used to focus the acoustic beam for defect detection. The focal length F of the focused ultrasonic transducer can, for example, satisfy: , where r is the radius of curvature of the acoustic lens, is the longitudinal wave velocity in the acoustic lens, is the longitudinal wave velocity in water. Further, during the immersion test using a focused ultrasonic transducer, the focal length satisfy: , where L is the distance from the focal position to the metal surface to be detected, The longitudinal velocity of the ultrasonic wave in the metal being tested. During testing, the metal being tested is placed in a water tank with the ultrasonic transducer positioned directly above it, with the acoustic beam axis incident vertically. The transducer's focal depth is varied, and the optimal focal depth is determined to be the focal depth at which the ultrasonic signal's amplitude reaches its maximum at the moment of testing.

[0094] In actual implementation scenarios, the peak-to-peak values ​​of the reflection signals generated by metals of different materials and defects of different depths are different. In one implementation scenario, the peak-to-peak value of the reflection wave of the surface defect of iron material is 1.4502, and the peak-to-peak value of the reflection wave of the surface defect of aluminum oxide material is 1.3715. In another implementation scenario, iron material is used as the metal body and aluminum oxide material is used as the defect. For example, when the aluminum oxide defect is 2 meters away from the upper surface of the iron body, the peak-to-peak value of the reflection wave of the surface defect of aluminum oxide material is 1.3715. In this case, the peak-to-peak value of the reflected wave is 1.4795. When the alumina defect is 6 In this case, the peak-to-peak value of the reflected wave is 1.4945.

[0095] In view of this, before executing step S101, simulation can be used to obtain the peak-to-peak value range, transmission time range, etc. of the reflection signal of the surface defect of the metal material to determine the threshold value of the signal intensity of the reflection wave signal, as well as the correspondence between the signal intensity of the reflection wave signal and the defect depth. In this way, an image can be generated by the peak position of the reflection wave signal, which is conducive to distinguishing the material of the metal to be detected and defects of different depths.

[0096] In some embodiments, before acquiring an image of the metal surface to be identified using an ultrasonic signal of a target frequency, a pre-deployed simulation model may be called in response to a creation instruction to create a metal specimen model and determine specimen parameters of the metal specimen model, wherein the specimen parameters include the location and size of surface defects in the metal specimen model. Simulation parameters may then be received, wherein the simulation parameters include the frequency of the ultrasonic signal, the ultrasonic signal transmission boundary of the metal specimen model, and grid parameters, wherein the grid parameters are used to divide the metal specimen model into a plurality of grid cells, and the maximum time step corresponding to the simulation model is calculated based on the grid parameters. Furthermore, the simulation model may be called to perform ultrasonic microscopy simulation on the metal specimen model based on the simulation parameters to determine the corresponding relationship between the signal intensity of the reflected wave signal and the defect depth. Subsequently, the corresponding relationship between the signal intensity of the reflected wave signal and the grayscale value may be determined based on the corresponding relationship between the signal intensity of the reflected wave signal and the defect depth.

[0097] For example, taking the use of the COMSOL Multiphysics model for simulation as an example, you can respond to the creation instruction and build a two-dimensional rectangular iron-based specimen in the COMSOL Multiphysics model, determine the length of the specimen to be, for example, 0.8 mm, determine the length of the specimen to be, for example, 0.5 mm, and set a defect with a side length of 0.05 mm at the center of the upper surface of the specimen. Among them, the COMSOL Multiphysics model is a finite element simulation model. In order to better study the effective information of the defect reflection signal and reduce the impact of the lateral propagation of the ultrasonic signal on the signal, the boundaries on both sides of the iron-based specimen can be set as low-reflection boundaries, and fixed constraints can be set for the bottom and top surfaces of the iron-based specimen. In this way, the defect echo interference caused by the reflected echo at the boundary of the iron-based specimen can be reduced.

[0098] In order to ensure the accuracy of calculation and simplify the amount of calculation, some optional designs can use free-style triangular mesh unit division, the maximum unit of the mesh Ultrasonic wavelength can be used 1 / 6 of the side length, that is, . Furthermore, the largest unit of the grid The maximum time step can be obtained , where the maximum time step is Can satisfy: ,in, is the longitudinal wave velocity. In an optional example, the maximum unit of the liquid grid can be calculated They are approximately 0.0015 mm, the largest unit of the grid in the solid The maximum time step is 0.004 mm. 0.0003 .

[0099] After calculating the above simulation parameters, an ultrasonic transducer with a pulse excitation frequency of 300 MHz can be selected to simulate the iron-based sample, thereby simulating the corresponding relationship between the signal intensity of the reflected wave signal and the defect depth. Then, based on the corresponding relationship between the signal intensity of the reflected wave signal and the defect depth, the corresponding relationship between the signal intensity of the reflected wave signal and the grayscale value is determined. In this way, when using the ultrasonic microscopy detection equipment for image acquisition, the image is generated based on the peak position of the reflected wave signal.

[0100] Step S102 : extracting at least one feature to be identified from the image, wherein any feature to be identified includes defect features of the metal surface to be identified.

[0101] In some embodiments, when the at least one feature to be identified includes two or more features to be identified, the scales of defect features included in any two features to be identified are different.

[0102] Combined with the above description of the recognition system, it can be seen that after the computer device collects the image of the metal surface to be identified through the ultrasonic microscopy detection equipment, the image can be input into the defect recognition model so that the defect recognition model processes the image and outputs the defect recognition result.

[0103] In some embodiments, the computer device extracts at least one feature to be identified from the image, which may include: extracting high-level features and low-level features of the image respectively, the high-level features including high-level features of defects on the metal surface to be identified, and the low-level features including low-level features of defects on the metal surface to be identified; splicing the high-level features and the low-level features to obtain image features of the metal surface to be identified; and extracting the at least one feature to be identified from the image features.

[0104] Among them, the high-level features of defects can include complex features such as defect shape features and category features. The low-level features of defects include features such as defect contour features and pixel values.

[0105] To facilitate understanding, the following describes the implementation of defect recognition in conjunction with the structure of the defect recognition model.

[0106] See also Figure 3A , Figure 3A An exemplary defect recognition model is shown. The model includes, according to data flow, a preprocessing network, a backbone network, a neck network, and a prediction network. In some embodiments, the backbone network may include a first feature extraction network and a second feature extraction network. The first feature extraction network includes a first number of cascaded feature extraction modules, and the second feature extraction network includes a second number of feature extraction modules.

[0107] After acquiring an image of the metal surface to be identified, the computer device can call the preprocessing network of the defect recognition model to reduce noise on the image, scale the image size, etc. The scaled image size is, for example, 992×992×3, that is, the width and height are both 992 pixels, and the color values ​​of 3 channels. After scaling the image to a size that matches the defect recognition model, the scaled image is transferred to the backbone network. Furthermore, the first feature extraction network in the backbone network can be called to extract high-level features of the image, and the second feature extraction network in the backbone network can be called to extract low-level features of the image. Afterwards, the high-level features and the low-level features are spliced ​​together to obtain the image features of the metal surface to be identified.

[0108] The feature extraction modules in the first feature extraction network and the second feature extraction network can be convolution modules. Any feature extraction module is, for example, a convolution-batch normalization-Leaky ReLU activation function (CBL) module. Each CBL module includes a convolution layer, a batch normalization layer, and an activation function layer. The convolution layer is used to extract features, the batch normalization layer is used to normalize feature data to accelerate training and improve stability, and the activation function avoids neuron death by nonlinearly processing feature data.

[0109] It should be understood that for the same image, features extracted by performing more convolutions can represent higher-level semantics and more complex representations than features extracted by performing fewer convolutions. Based on this, the first number is greater than the second number, that is, the number of feature extraction modules included in the first feature extraction network is greater than the number of feature extraction modules included in the second feature extraction network.

[0110] For example, Figure 3A A network architecture of the backbone network can be as follows Figure 3B As shown, Figure 3B The first feature extraction network in the backbone network contains five CBL modules, while the second feature extraction network contains one CBL module. In this way, the first feature extraction network performs five convolutions on the image of the metal surface to be identified, extracting higher-level features from the image and thus the high-level features of the metal surface defects to be identified. The second feature extraction network performs one convolution on the image of the metal surface to be identified, extracting lower-level features from the image and thus the low-level features of the metal surface defects to be identified.

[0111] based on Figure 3BIt can be seen that the first feature extraction network and the second feature extraction network form a residual structure. Therefore, splicing high-level features and low-level features together as image features is conducive to obtaining features of different levels of metal surface defects to be identified, thereby improving the accuracy of identifying micron-level defects.

[0112] See again Figure 3B In some embodiments, according to the flow of data, after the three cascaded CBL modules of the first feature extraction network, four parallel pooling layers can be connected, and the four pooling layers are connected to two cascaded CBL modules. The pooling kernels of the four pooling layers are different, and the pooling kernels of the four pooling layers are, for example, 1×1, 5×5, 9×9, and 13×13. In this way, after the first feature extraction network extracts features from the image through three cascaded CBL modules, it can pool the proposed features from four dimensions, and then splice the pooled features, and then undergo two more feature extractions to obtain high-level features of the image. The adoption of this implementation method is conducive to alleviating the gradient vanishing problem caused by the increase in network depth and enhancing the expression ability of the backbone network.

[0113] Further, see again Figure 3A After the backbone network extracts the image features, the image features can be input into the neck network. The neck network extracts the at least one feature to be identified from the image features, and then inputs the extracted at least one feature to be identified into the prediction network.

[0114] In some embodiments, if one feature to be identified is extracted from the image features, the neck network may deploy a convolution module to extract the feature to be identified from the image features. In other embodiments, if at least two features to be identified are extracted from the image features, the neck network may deploy at least two convolution modules, each having a different convolution kernel and number of channels, and each of the at least two convolution modules extracts one feature to be identified from the image features.

[0115] Exemplarily, when the at least one feature to be identified includes two or more features to be identified, the neck network may deploy a Feature Pyramid Network (FPN), for example. Using the FPN to extract the at least one feature to be identified from the image features may include: upsampling the image features N times to obtain N enhanced features of the image features, downsampling the enhanced features obtained by the N-th upsampling N times to obtain N reduced-dimensionality features, and, for any reduced-dimensionality feature, concatenating the reduced-dimensionality feature and the enhanced feature having the same scale as the reduced-dimensionality feature to obtain the feature to be identified. N is an integer greater than or equal to 2.

[0116] The FPN structure is as follows Figure 3C As shown, accordingly, three features to be identified can be extracted, and the three features to be identified can respectively include large-scale features, medium-scale features and small-scale features of the defect.

[0117] It can be seen that by adopting this implementation method, through multiple upsampling and downsampling, features with different spatial resolutions and rich semantic information can be obtained, so that characterization data of defect features of different scales can be obtained, which is conducive to improving the recognition ability of micron-level defects of different sizes.

[0118] It should be understood that the classification of defects into large, medium, and small sizes in the embodiments of the present application can be flexible and based on the size distribution of micron-scale defects on metal surfaces in actual implementation scenarios, rather than a fixed data range. For example, if the size distribution of micron-scale defects on a metal surface ranges from 3 microns to 50 microns, then defects less than or equal to 10 microns can be classified as small, defects greater than 10 microns to less than or equal to 25 microns can be classified as medium, and defects greater than 25 microns can be classified as large.

[0119] Step S103 : Identify the at least one feature to be identified respectively to obtain at least one defect identification sub-result.

[0120] The at least one feature to be identified corresponds one-to-one to the at least one defect identification sub-result.

[0121] Combined with the description of the features to be identified in the above embodiment, it can be seen that in the case where at least one feature to be identified includes two or more features to be identified, the scales of the defect features contained in each feature to be identified are different. In order to ensure the recognition effect of defect features of different scales, for the features to be identified that include defect features of different scales, recognition modules that match the corresponding scales can be deployed to perform defect recognition. Correspondingly, in the scenario where at least one feature to be identified includes two or more features to be identified, Figure 3A The prediction network shown in the figure can deploy two or more recognition modules, and the number of recognition modules can be the same as the number of features to be recognized.

[0122] In some embodiments, any recognition module may include convolutional layers, pooling layers, and fully connected layers. The scales of the convolutional layers and pooling layers of different recognition modules vary and can match the scale of the features to be recognized. Based on this, for any feature to be recognized, a recognition module that matches the scale of the feature to be recognized can be called to identify defect features in the feature to be recognized, thereby obtaining a defect recognition sub-result corresponding to the feature to be recognized.

[0123] The following describes the neck network, the scale of the features to be identified, the prediction network, and the recognition module by taking the example of at least one feature to be identified including three features to be identified.

[0124] like Figure 3C As shown, taking the neck network implemented as FPN as an example, after receiving the image features, FPN sequentially upsamples the image features three times according to three cascaded upsampling convolution layers, and each upsampling convolution layer can output an enhanced feature of the image features. The enhanced features sampled by the third upsampling convolution layer are used as the input features of the three cascaded downsampling convolution layers. Then, the enhanced features are sequentially downsampled three times, and each downsampling convolution layer can output a reduced-dimensional feature of the enhanced features. Among them, the upsampling resolution order of the three upsampling convolution layers is opposite to the downsampling resolution order of the three downsampling convolution layers. The reduced-dimensional features output by each downsampling convolution layer are spliced ​​with the enhanced features output by the corresponding upsampling convolution layer. The spliced ​​features are the features to be identified corresponding to the downsampling convolution layer. The corresponding upsampling convolution layer refers to the upsampling convolution layer with the same resolution as the downsampling convolution layer.

[0125] according to Figure 3C As shown, the three downsampling convolution layers of the FPN are connected one-to-one to the three recognition modules of the prediction network. The three features to be recognized are, for example, features to be recognized at a first scale, features to be recognized at a second scale, and features to be recognized at a third scale. The first scale is larger than the second scale, and the second scale is larger than the third scale. According to the data flow, the first downsampling convolution layer in the FPN transmits the features to be recognized at the first scale to the first recognition module, the second downsampling convolution layer in the FPN transmits the features to be recognized at the second scale to the second recognition module, and the third downsampling convolution layer in the FPN transmits the features to be recognized at the third scale to the third recognition module. Furthermore, the first recognition module can be called to identify the features to be recognized at the first scale to obtain the defect recognition sub-result of the first scale of the metal surface to be recognized, the second recognition module can be called to identify the features to be recognized at the second scale to obtain the defect recognition sub-result of the second scale of the metal surface to be recognized, and the third recognition module can be called to identify the features to be recognized at the third scale to obtain the defect recognition sub-result of the third scale of the metal surface to be recognized.

[0126] For example, according to the number of convolution layers × the number of convolution kernels × the number of channels, the scale of the first recognition module can be, for example, 124×124×256, the scale of the second recognition module can be, for example, 62×62×512, and the scale of the third recognition module can be, for example, 31×31×1024.

[0127] It should be understood that the above Figure 3CThis is merely a schematic description and does not limit the architecture of the neck network and prediction network in the embodiments of the present application. In actual implementation scenarios, the number of upsampling convolutional layers or downsampling convolutional layers included in the neck network, as well as the number of recognition modules included in the prediction network, can be greater or lesser. In addition, the scale of each recognition module in the prediction network can also be other values. The embodiments of the present application do not limit this.

[0128] Step S104: combining the at least one defect recognition sub-result to obtain a defect recognition result of the metal surface to be recognized.

[0129] The defect identification result includes at least one of the defect type, defect location, and defect size. For example, the defect type may include cracks, holes, inclusions, protrusions, etc.

[0130] In combination with the above description of the features to be identified and the identification modules, it can be seen that when at least one feature to be identified is implemented as one feature to be identified, a recognition module included in the prediction network outputs a defect identification sub-result, and the defect identification sub-result is the defect identification result of the metal surface to be identified. When at least one feature to be identified is implemented as at least two features to be identified, the at least two recognition modules included in the prediction network respectively output defect identification sub-results, and the defects included in different defect identification sub-results can be defects of different sizes in the metal surface to be identified. For example, the defect identification sub-result of the first scale is a defect less than or equal to 10 microns, the defect identification sub-result of the second scale is a defect greater than 10 microns and less than or equal to 25 microns, and the defect identification sub-result of the third scale is a defect greater than 25 microns. Combining the various defect identification sub-results can obtain the defect identification result of the metal surface to be identified.

[0131] It can be seen that in the scenario of outputting at least two defect identification sub-results, the embodiment of the present application can identify defects of different sizes by deploying deep learning networks with different parameters, which is conducive to the comprehensive and accurate identification of micron-level defects on the metal surface to be identified.

[0132] In some embodiments, before step S102, a training sample set may be generated and used to train a deep learning model to obtain the above-mentioned defect recognition model. Each training sample includes a label, and the label is used to indicate whether the training sample contains a surface defect and, if so, the location, size, and category of the surface defect.

[0133] In actual implementation scenarios, multiple metal samples can be scanned with an ultrasonic signal at the target frequency to obtain multiple initial training sample images. These initial training sample images can be cropped from multiple metal sample images, with the size of the cropped initial training sample images meeting the input image size of the deep learning model. In response to a labeling instruction, for any initial training sample image containing surface defects, labels are added to each surface defect in the initial training sample image. Each label is used to annotate the location, size, and category of the corresponding surface defect.

[0134] For example, a type identifier can be pre-set for each defect type. For example, a crack is identified as 01, a cavity is identified as 10, a protrusion is identified as 11, and a non-defect is identified as 00. Furthermore, a corresponding defect category identifier can be added to any initial training sample image. For defects contained in the initial training sample image, a rectangular annotation box can be marked in the defect area. The location and size of the defect can be identified by setting the horizontal coordinate, vertical coordinate, width, and height of the center point of the rectangular annotation box.

[0135] In order to expand the data of the training samples, data augmentation processing is performed on the training samples. Every M labeled initial training sample images can be spliced ​​into one training sample to obtain the training sample set, where M is an integer greater than or equal to 2. For example, every 4 initial training sample images can be spliced ​​into one training sample.

[0136] Furthermore, the deep learning model can be invoked to predict the input training samples, outputting the predicted location, predicted size, and predicted category of the surface defects in the training samples. A loss function is then calculated. If the loss function meets the convergence criteria, the deep learning model is determined as the defect recognition model. If the loss function does not meet the convergence criteria, a new training sample is determined from the training sample set, and the step of invoking the deep learning model to predict the input training sample is repeated for the new training sample.

[0137] The loss function is obtained by weighted summation of positioning loss, classification loss and confidence loss, that is, the loss function satisfy: , is the positioning loss, is the classification loss, Refers to the confidence loss. The positioning loss Characterize the loss between the predicted position and the position marked by the training sample, the classification loss Characterize the loss between the predicted category and the category labeled by the training sample, the confidence loss Characterize the cross entropy loss between the predicted category and the category labeled by the training sample.

[0138] Positioning loss satisfy: , where IoU is the ratio of the intersection and union of the predicted defect area box and the annotation box, and the calculation formula is: ,in, is the area box of the predicted defect, For the label box, is the center point of the predicted defect area box, The center point of the annotation box, is the Euclidean distance, is the maximum distance between the predicted defect area box and the annotation box boundary, is the weight coefficient, , is the width of the predicted defect area box, is the height of the predicted defect area box, is the width of the annotation box, The height of the dimension box.

[0139] It can be seen that the positioning loss Not only the overlap area (IoU) between the predicted position and the annotated position is considered, but also the shape and size of the predicted position and the annotated position are considered, so that the model can more accurately regress the position and size of the annotated box, thereby improving the accuracy of defect recognition.

[0140] Classification loss satisfy: ,in, Refers to the defect category marked, is the predicted category.

[0141] It can be seen that by comprehensively considering the positioning, classification and confidence dimensions of the model, the defect recognition model can have better performance and robustness in the defect recognition task.

[0142] The recognition system of the embodiment of the present application uses an ultrasonic signal with a target frequency greater than a preset frequency threshold to capture images of the metal surface to be identified. Since the higher the frequency of the ultrasonic signal, the stronger the ability to detect micron-level defects, based on this, the image captured using a higher-frequency ultrasonic signal can, to a certain extent, improve the detection rate of smaller micron-level defects, so that the captured image can more clearly show the characteristics of smaller defects. Furthermore, the recognition system extracts the features of the image and identifies the defects of the metal surface to be identified based on the features of the image. Not only does it not require manual identification, but it also uses a deep learning model to identify defects, which can improve the accuracy of identification and reduce the missed detection rate.

[0143] In order to intuitively demonstrate the defect recognition effect of the technical solution of this application, see Figure 4A A comparison chart showing defect recognition results using different recognition methods. Figure 4A (a) is the image of the metal surface to be identified obtained by ultrasonic scanning at a conventional frequency. Figure 4A (b) is the recognition module that comes with conventional ultrasonic microscopy equipment. Figure 4A The recognition result of (a) in Figure 4A (c) is correct Figure 4A The manual recognition result of (a) in Figure 4A (d) is the defect recognition model of this application. Figure 4A The recognition result of (a) in the figure is shown. Clearly, since the metal surface cannot be guaranteed to be absolutely flat, the scanned image quality is poor, with varying brightness levels. The built-in recognition module of the ultrasonic microscope only uses simple threshold adjustment to identify defects, which is ineffective. Manual labeling relies heavily on manual experience and is difficult to meet the efficiency and high-precision requirements of actual production. The deep learning model used in this application not only improves recognition efficiency for low-quality images, but also helps improve the accuracy of defect identification and reduce missed detection rates.

[0144] Figure 4B The image comparison diagram of the metal surface to be identified is shown. Figure 4B (a) is an image acquired by using the ultrasonic microscopy technology of the embodiment of the present application. Figure 4B Image (b) is captured using an electron microscope. The imaging results show that the number of defects detected using the ultrasonic microscopy technique in this embodiment of the present application is higher than that detected using an electron microscope, while the dimensional accuracy of the defects detected using the ultrasonic microscopy technique in this embodiment of the present application is significantly superior to that detected using an electron microscope. This shows that ultrasonic microscopy, with its penetrating power, can effectively detect superficial defects, fully demonstrating the advantages of ultrasonic microscopy in material defect detection.

[0145] It should be noted that:

[0146] In the description provided herein, a large number of specific details are described. However, it is understood that the embodiments of the present application can be practiced without these specific details. In some instances, well-known structures and technologies are not shown in detail so as not to obscure the understanding of this description.

[0147] Similarly, it should be understood that in order to streamline the present application and aid in understanding one or more of the various inventive aspects, in the above description of the exemplary embodiments of the present application, various features of the present application are sometimes grouped together into a single embodiment, figure, or description thereof. However, the above-claimed application requires more features than those explicitly recited in each claim. More specifically, as reflected in the claims below, inventive aspects lie in fewer than all the features of the individual embodiments disclosed above. Therefore, the claims following the detailed description are hereby expressly incorporated into this detailed description, with each claim itself serving as a separate embodiment of the present application.

[0148] Furthermore, those skilled in the art will appreciate that although some embodiments described herein include certain features included in other embodiments but not other features, combinations of features from different embodiments are intended to be within the scope of this application and to form different embodiments. For example, in the claims below, any of the claimed embodiments may be used in any combination.

[0149] The above description is merely a preferred embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for identifying metal surface micro-defects based on ultrasonic microscopy, characterized in that: Applied to an identification system, the method includes: Acquiring an image of the metal surface to be identified using an ultrasonic signal of a target frequency, wherein the target frequency is 300 MHz; extracting two or more features to be identified from the image, wherein any one of the features to be identified includes a defect feature of the metal surface to be identified, and the scales of the defect features included in any two features to be identified are different; Respectively identifying the two or more features to be identified to obtain two or more defect identification sub-results, wherein the two or more features to be identified correspond one to one with the two or more defect identification sub-results; Combining the two or more defect identification sub-results to obtain a defect identification result of the metal surface to be identified, wherein the defect identification result includes at least one of a defect type, a defect location, and a defect size; The step of extracting two or more features to be identified from the image includes: Extracting high-level features and low-level features of the image respectively, the high-level features including shape features and category features of the defects on the metal surface to be identified, and the low-level features including contour features and pixel values ​​of the defects on the metal surface to be identified; splicing the high-level features and the low-level features to obtain image features of the metal surface to be identified; Extracting the two or more features to be identified from the image features; The extracting high-level features and low-level features of the image respectively includes: Invoking a first feature extraction network to extract high-level features of the image, the first feature extraction network comprising a first number of cascaded feature extraction modules; Invoking a second feature extraction network to extract low-level features of the image, the second feature extraction network including a second number of feature extraction modules, the first number being greater than the second number; The extracting the two or more features to be identified from the image features includes: Upsampling the image feature N times to obtain N enhanced features of the image feature, where N is an integer greater than or equal to 2; The enhanced features obtained by the Nth upsampling are downsampled N times to obtain N dimensionality reduction features respectively; For any dimensionality reduction feature, the dimensionality reduction feature and the enhanced feature with the same scale as the dimensionality reduction feature are concatenated to obtain a feature to be identified; The step of respectively identifying the two or more features to be identified to obtain two or more defect identification sub-results includes: For any feature to be identified, calling a recognition module that matches the scale of the feature to be identified, identifying defect features in the feature to be identified, and obtaining a defect identification sub-result corresponding to the feature to be identified; The number of recognition modules is the same as the number of features to be recognized. The scales of the convolutional layers and pooling layers of different recognition modules are different and match the scales of the features to be recognized.

2. The method for identifying metal surface micro-defects based on ultrasonic microscopy according to claim 1, characterized in that: The method of collecting an image of the metal surface to be identified using an ultrasonic signal of a target frequency includes: emitting an ultrasonic signal at a frequency of 300 MHz toward the metal surface to be identified according to a scanning trajectory matching the metal surface to be identified; For any scanning position of the metal surface to be identified, receiving a reflected wave signal of the scanning position to the ultrasonic signal; According to the preset correspondence between the signal intensity and the grayscale value of the reflected wave signal, the grayscale value of the imaging pixel corresponding to the scanning position is determined to obtain the image of the metal surface to be identified.

3. The method for identifying metal surface micro-defects based on ultrasonic microscopy according to claim 1 or 2, characterized in that: Before acquiring an image of the metal surface to be identified using an ultrasonic signal of a target frequency, the method further includes: In response to the creation instruction, calling a pre-deployed simulation model to create a metal sample model, and determining sample parameters of the metal sample model, wherein the sample parameters include positions and sizes of surface defects in the metal sample model; receiving simulation parameters, the simulation parameters including the frequency of the ultrasonic signal, the ultrasonic signal transmission boundary of the metal sample model, and grid parameters, the grid parameters being used to divide the metal sample model into a plurality of grid units; Calculating the maximum time step corresponding to the simulation model according to the grid parameters; Calling the simulation model to perform ultrasonic microscopy simulation on the metal sample model based on the simulation parameters to determine a corresponding relationship between the signal intensity of the reflected wave signal and the defect depth; According to the corresponding relationship between the signal intensity of the reflected wave signal and the defect depth, the corresponding relationship between the signal intensity of the reflected wave signal and the grayscale value is determined.

4. The method for identifying metal surface micro-defects based on ultrasonic microscopy according to claim 1, characterized in that: The two or more features to be identified include features to be identified at a first scale, features to be identified at a second scale, and features to be identified at a third scale, wherein the first scale is larger than the second scale, and the second scale is larger than the third scale; The step of calling a recognition module that matches the scale of any feature to be recognized to recognize the defect feature in the feature to be recognized includes: Invoking a first recognition module to recognize the feature to be recognized at the first scale, and obtaining a defect recognition sub-result of the first scale of the metal surface to be recognized; Invoking a second recognition module to recognize the feature to be recognized at the second scale, and obtaining a defect recognition sub-result of the second scale of the metal surface to be recognized; The third recognition module is called to recognize the feature to be recognized at the third scale, and a defect recognition sub-result of the third scale of the metal surface to be recognized is obtained.

5. The method for identifying metal surface micro-defects based on ultrasonic microscopy according to claim 1, characterized in that: Before extracting two or more features to be identified from the image, the method further includes: Generating a training sample set, wherein each training sample includes a label, wherein the label is used to mark whether the training sample contains a surface defect, and if the training sample contains a surface defect, mark the position, size, and category of the surface defect; The training sample set is used to train a deep learning model to obtain a defect recognition model, and the defect recognition model is used to input the image and output the defect recognition result.

6. The method for identifying metal surface micro-defects based on ultrasonic microscopy according to claim 5, characterized in that: Generating a training sample set includes: Scanning a plurality of metal samples with an ultrasonic signal of the target frequency to obtain a plurality of initial training sample images; For any initial training sample image with surface defects, in response to a labeling instruction, a label is added to each surface defect of the initial training sample image, where each label is used to mark the position, size, and category of the corresponding surface defect; Each M labeled initial training sample images are spliced ​​into a training sample to obtain the training sample set, where M is an integer greater than or equal to 2.

7. The method for identifying metal surface micro-defects based on ultrasonic microscopy according to claim 5 or 6, characterized in that: The deep learning model is trained using the training sample set to obtain a defect recognition model, including: Calling the deep learning model to predict the input training sample, and outputting the predicted position, predicted size and predicted category of the surface defect of the training sample; Calculating a loss function, the loss function being a weighted sum of positioning loss, classification loss, and confidence loss; the positioning loss representing the loss between the predicted position and the position labeled with the training sample, the classification loss representing the loss between the predicted category and the category labeled with the training sample, and the confidence loss representing the cross entropy loss between the predicted category and the category labeled with the training sample; If the loss function satisfies the convergence condition, determining the deep learning model as the defect recognition model; If the loss function does not meet the convergence condition, a new training sample is determined from the training sample set, and the step of calling the deep learning model to predict the input training sample is performed again for the new training sample.

8. A metal surface micro-defect recognition system based on ultrasonic microscopy, characterized in that: The identification system includes a computer device and an ultrasonic microscopy detection device, wherein a defect recognition model is deployed in the computer device, wherein: The ultrasonic microscopy detection device is configured to receive a trigger instruction from the computer device, capture an image of the metal surface to be identified using an ultrasonic signal at a target frequency, and transmit the image to the computer device, wherein the target frequency is 300 MHz; The computer device is configured to call the defect recognition model to extract two or more features to be recognized from the image, wherein any of the features to be recognized includes a defect feature of the metal surface to be recognized, and the scales of the defect features included in any two features to be recognized are different; respectively recognize the two or more features to be recognized to obtain two or more defect recognition sub-results, wherein the two or more features to be recognized correspond one-to-one to the two or more defect recognition sub-results; and combine the two or more defect recognition sub-results to obtain a defect recognition result of the metal surface to be recognized, wherein the defect recognition result includes at least one of a defect type, a defect location, and a defect size; The step of extracting two or more features to be identified from the image includes: Extracting high-level features and low-level features of the image respectively, the high-level features including shape features and category features of the defects on the metal surface to be identified, and the low-level features including contour features and pixel values ​​of the defects on the metal surface to be identified; splicing the high-level features and the low-level features to obtain image features of the metal surface to be identified; Extracting the two or more features to be identified from the image features; The extracting high-level features and low-level features of the image respectively includes: Invoking a first feature extraction network to extract high-level features of the image, the first feature extraction network comprising a first number of cascaded feature extraction modules; Invoking a second feature extraction network to extract low-level features of the image, the second feature extraction network including a second number of feature extraction modules, the first number being greater than the second number; The extracting the two or more features to be identified from the image features includes: Upsampling the image feature N times to obtain N enhanced features of the image feature, where N is an integer greater than or equal to 2; The enhanced features obtained by the Nth upsampling are downsampled N times to obtain N dimensionality reduction features respectively; For any dimensionality reduction feature, the dimensionality reduction feature and the enhanced feature with the same scale as the dimensionality reduction feature are concatenated to obtain a feature to be identified; The step of respectively identifying the two or more features to be identified to obtain two or more defect identification sub-results includes: For any feature to be identified, calling a recognition module that matches the scale of the feature to be identified, identifying defect features in the feature to be identified, and obtaining a defect identification sub-result corresponding to the feature to be identified; The number of recognition modules is the same as the number of features to be recognized. The scales of the convolutional layers and pooling layers of different recognition modules are different and match the scales of the features to be recognized.

9. The metal surface micro-defect identification system based on ultrasonic microscopy according to claim 8, characterized in that: The ultrasonic microscopic detection equipment includes a pulse transceiver, a scanner, an ultrasonic transducer and a data collector, wherein: The scanner is configured to respond to the trigger instruction and move the ultrasonic transducer along a scanning trajectory that matches the metal surface to be identified; The pulse transceiver is used to send an excitation signal to the ultrasonic transducer in response to the trigger instruction; The ultrasonic transducer is configured to transmit an ultrasonic signal at a frequency of 300 MHz to the metal surface to be identified in response to the excitation signal; and is further configured to receive a reflected wave signal of the ultrasonic signal at any scanning position on the metal surface to be identified; The data collector is used to determine the grayscale value of the imaging pixel corresponding to the scanning position according to the preset corresponding relationship between the signal intensity and the grayscale value of the reflected wave signal, so as to obtain the image of the metal surface to be identified.

Images