A magnetic material defect detection system based on image recognition

Abstract

The application discloses a kind of magnetic material defect detection systems based on image recognition, using first acquisition module, first determination module, second determination module, second acquisition module and third acquisition module, first acquisition module is used to adopt polarized light imaging device to carry out multi-angle light illumination control to the surface of magnetic material, by adjusting polaroid angle and light source incidence angle, specular reflection generated by suppressing strong light characteristic, obtain the original image data with uniform illumination distribution, if it is detected that there is overexposure area in image, then automatically adjust exposure parameter and polarization angle, obtain the high-quality image suitable for subsequent processing;First determination module is used to carry out multi-scale wavelet transform decomposition according to the high-quality image obtained, identify surface defect area by analyzing the texture features and edge information in different frequency components, if wavelet coefficient exceeds preset threshold in high-frequency component, then it is judged as potential defect area.The detection performance of the application is significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of magnetic material defect detection, and discloses a magnetic material defect detection system based on image recognition. BACKGROUND

[0002] As a core basic material of modern industry, magnetic materials are widely used in key fields such as motors, transformers, and nuclear power equipment, and the quality of the magnetic materials directly affects the safety and reliability of the entire system. With the increasing demand for product precision in high-end manufacturing, defect detection of magnetic materials has become an important link to ensure product quality.

[0003] Traditional magnetic material defect detection mainly relies on manual visual inspection and simple magnetic powder detection methods, which have obvious technical limitations. Manual detection is not only inefficient and time-consuming, but also affected by the experience and subjective judgment of the detector, resulting in a high misjudgment rate.

[0004] Although existing automated detection equipment has improved detection efficiency to some extent, it still has problems such as insufficient detection accuracy and poor adaptability when facing complex magnetic material defect types.

[0005] The core challenge of magnetic material defect detection is due to the complex characteristics of the material itself. The strong light reflection characteristics of the surface of the magnetic material make it difficult for optical imaging systems to obtain clear and stable images. This instability of imaging quality directly affects the accuracy of subsequent defect recognition. When the imaging system cannot provide high-quality basic data, traditional image processing algorithms cannot accurately distinguish between real defects and imaging noise, resulting in a serious lack of recognition ability for small defects in the detection system.

[0006] More complex is that the defects of magnetic materials often present diversified characteristics, including surface visible cracks and edge collapse, as well as internal hidden pores and inclusions. The diversity of defect types requires the detection system to have multi-modal information fusion processing capability, but the existing technology is obviously insufficient in this regard.

[0007] Therefore, how to realize high-quality image acquisition in a strong light interference environment and establish an intelligent detection system that can simultaneously recognize multiple surface and internal defect types has become a key problem in the development of magnetic material defect detection technology. SUMMARY

[0008] The present application provides a magnetic material defect detection system based on image recognition, which aims to solve at least one of the defects in the prior art.

[0009] The present application relates to a magnetic material defect detection system based on image recognition, comprising:

[0010] The first acquisition module is configured to control multi-angle light irradiation on the surface of the magnetic material by using a polarized light imaging device, suppress mirror reflection caused by strong light reflection characteristics by adjusting the angle of a polarizer and the incident angle of a light source, acquire original image data with uniform light distribution, and automatically adjust exposure parameters and the polarization angle if overexposed areas are detected in the image to obtain high-quality images suitable for subsequent processing.

[0011] The first determination module is configured to perform multi-scale wavelet transform decomposition based on the acquired high-quality images, identify surface defect areas by analyzing texture features and edge information in different frequency components, and determine defect boundary contours and geometric feature parameters by combining gradient information if wavelet coefficients exceed a preset threshold in high-frequency components.

[0012] The second determination module is configured to scan the internal structure of the magnetic material by using an eddy current detection sensor array, acquire amplitude and phase change data of electromagnetic induction signals, determine the presence of internal defects if the signal amplitude change rate exceeds a reference value, and analyze phase differences by using a signal processing algorithm to determine the depth position and size range of internal defects.

[0013] The second acquisition module is configured to construct a multi-dimensional feature vector based on the geometric feature parameters of surface defects and the electromagnetic signal characteristics of internal defects, train a defect classification model by using a support vector machine algorithm, determine whether the defect type belongs to a crack, a pore, or an inclusion by using the spatial distribution pattern of the feature vector, and obtain a confidence score for each defect type.

[0014] The third acquisition module is configured to integrate surface optical detection results and internal electromagnetic detection results by using a weighted fusion algorithm, improve the defect confidence of a region if defects are identified in the region by both surface optical detection and internal electromagnetic detection, determine a final defect detection report by using spatial position matching and feature similarity calculation, and obtain a comprehensive evaluation result including defect position, type, and severity.

[0015] Further, the first acquisition module includes:

[0016] The first acquisition unit is configured to perform multi-angle light scanning on the surface of the magnetic material by adjusting the angle of a polarizer and the incident angle of a light source based on the characteristics of polarized light imaging, acquire a first set of image data from the scanning results, perform brightness distribution detection on the first set of image data, determine whether overexposed areas exist, and obtain a preliminary brightness distribution evaluation result.

[0017] The first determining unit is configured to, if there is an overexposed area in the first group of image data, perform secondary light control on the overexposed area by automatically adjusting an exposure parameter and a polarizer angle, acquire a second group of image data under the adjusted light condition, and perform brightness uniformity detection on the second group of image data to determine whether there is a residual specular reflection area.

[0018] The judging unit is configured to, for the residual specular reflection area in the second group of image data, perform area segmentation on the residual specular reflection area by using a preset threshold, extract position information of a reflection interference area from a segmentation result, acquire a third group of image data by adjusting a local light source intensity and finely adjusting a polarization angle, and judge whether a uniform light distribution standard is reached.

[0019] The second determining unit is configured to, according to a uniformity detection result of the third group of image data, if there is still a local brightness uneven area, perform pixel-level correction on the local brightness uneven area by using an image processing tool, acquire a fourth group of image data from the corrected data, and determine whether a requirement for subsequent processing is met.

[0020] Further, the first determining module comprises:

[0021] The second acquiring unit is configured to, according to a principle of multi-scale decomposition, separate image data of different frequency components from the high-quality image by using a wavelet transform tool, extract a detail part in a high-frequency component to obtain initial decomposition data containing texture features and edge information.

[0022] The third acquiring unit is configured to, for the high-frequency component information in the initial decomposition data, compare the high-frequency component information with a preset threshold, if a wavelet coefficient of a certain area exceeds the preset threshold, determine that the area is a first potential defect area, and acquire specific position distribution data of the first potential defect area.

[0023] The extracting unit is configured to, according to the specific position distribution data of the first potential defect area, detect a change of pixels around the area by using a gradient information calculation tool, determine boundary contour information of the first potential defect area, and extract geometric feature data in the boundary contour.

[0024] The fourth acquiring unit is configured to, by using the boundary contour information and the geometric feature data, in combination with a distribution state of the texture features, classify and label the potential defect area by using an image processing tool to obtain a final defect recognition result.

[0025] Further, the second acquiring unit is specifically configured to separate image data of a low-frequency component and image data of a high-frequency component from the high-quality image by using a wavelet transform tool to perform layered processing on the high-quality image.

[0026] Further, the second determining module comprises:

[0027] The fifth acquisition unit is configured to comprehensively scan the internal structure of the material by using the sensor array layout, record electromagnetic induction signals by using a preset signal acquisition frequency, extract amplitude variation rate and phase difference value data from the electromagnetic induction signals, and obtain a preliminary signal feature distribution;

[0028] The third determination unit is configured to compare the amplitude variation rates point by point by using a signal processing tool for the preliminary signal feature distribution, determine that there is an internal defect in a region if the amplitude variation rate of the region exceeds a preset reference value, and determine the position coordinates of the second potential defect region.

[0029] The sixth acquisition unit is configured to analyze the depth of the internal defect layer by layer by using a depth positioning parameter according to the position coordinates of the second potential defect region and in combination with the phase difference value data, and obtain specific depth distribution information of the internal defect.

[0030] The fourth determination unit is configured to divide the defect size range by boundary by using a size estimation method for the specific depth distribution information of the internal defect, compare electromagnetic induction signals of adjacent regions by using a signal processing tool, and determine complete size range data of the internal defect.

[0031] Further, the second acquisition module includes:

[0032] The seventh acquisition unit is configured to perform edge detection on the defect region by using an image processing tool according to the surface geometric shape data, extract contour information of the defect from the collected surface image, and obtain a preliminary geometric feature set.

[0033] The fifth determination unit is configured to extract features of response signals of the internal defect by using a signal processing tool for the preliminary geometric feature set in combination with electromagnetic signal intensity data, construct a multi-dimensional feature vector containing surface and internal information, determine a comprehensive feature description result, and

[0034] The eighth acquisition unit is configured to divide the feature vectors by type by using a classification processing tool if the feature vectors in the comprehensive feature description result present a specific spatial distribution pattern, determine whether the defect belongs to a crack, a pore or an inclusion, and obtain classification determination data.

[0035] The ninth acquisition unit is configured to quantitatively process the confidence of each defect by using a probability calculation tool for specific features of the defect type according to the classification determination data, and obtain final score data.

[0036] Further, the third acquisition module includes:

[0037] The tenth acquisition unit is configured to scan the target region by using an optical imaging tool according to the data of surface detection, extract preliminary boundary information of the surface defect from a scanning image, and perform fine processing on the preliminary boundary information by using an edge detection tool to obtain accurate feature data of the surface defect.

[0038] The sixth determination unit is configured to perform signal collection on the target region by using an electromagnetic induction tool according to the requirement of internal detection, extract response features of the internal defect from the signal, and perform region matching on the response features and the accurate feature data of the surface defect by using a position calibration tool if the response features and the accurate feature data of the surface defect overlap in spatial position, to determine defect distribution records of the overlapping region.

[0039] The eleventh acquisition unit is configured to calculate the similarity degree of the surface defect feature and the internal defect feature by using a feature comparison tool according to the defect distribution records, and integrate the two detection results by using a weighted fusion tool if the similarity degree is higher than a preset threshold, to obtain a comprehensive defect feature set.

[0040] The generation unit is configured to perform type division on the defect by using a classification and discrimination tool according to the comprehensive defect feature set, and perform quantitative processing on the defect feature by using an intensity evaluation tool to determine the specific type and severity of the defect, and generate a defect detection report containing position information.

[0041] Further, in the sixth determination unit, the electromagnetic induction tool includes an eddy current detection sensor array.

[0042] The present application has the following beneficial effects:

[0043] The application discloses a magnetic material defect detection system based on image recognition, which adopts a first acquisition module, a first determination module, a second determination module, a second acquisition module and a third acquisition module. The first acquisition module is used for controlling multi-angle light irradiation of a polarized light imaging device on the surface of a magnetic material, suppressing specular reflection caused by strong light reflection characteristics by adjusting the angle of a polarizer and the angle of light source incidence, acquiring original image data with uniform light distribution, and automatically adjusting exposure parameters and the angle of polarization if overexposed areas are detected in the image to obtain high-quality images suitable for subsequent processing. The first determination module is used for multi-scale wavelet transform decomposition according to the acquired high-quality images, identifying surface defect areas by analyzing texture features and edge information in different frequency components, and determining defect boundary profiles and geometric feature parameters by combining gradient information if wavelet coefficients exceed a preset threshold in high-frequency components. The second determination module is used for scanning the internal structure of the magnetic material by an eddy current detection sensor array, acquiring amplitude and phase change data of electromagnetic induction signals, and judging that there is an internal defect if the signal amplitude change rate exceeds a reference value, and analyzing phase differences by a signal processing algorithm to determine the depth position and size range of the internal defect. The second acquisition module is used for constructing a multi-dimensional feature vector according to the geometric feature parameters of the surface defect and the electromagnetic signal features of the internal defect, training a defect classification model by a support vector machine algorithm, judging the defect type to be a crack, a pore or an inclusion by the spatial distribution mode of the feature vector, and obtaining a confidence score of each defect type. The third acquisition module is used for integrating surface optical detection results and internal electromagnetic detection results by a weighted fusion algorithm, improving the defect confidence of a region if defects are identified in the region by both surface optical detection and internal electromagnetic detection, and determining a final defect detection report by spatial position matching and feature similarity calculation to obtain a comprehensive evaluation result containing defect position, type and severity. The magnetic material defect detection system based on image recognition provided by the application realizes the leap of magnetic material defect detection from “single-mode local detection” to “full-dimensional intelligent diagnosis” through the triple superposition of the surface detail capturing capability of polarized light imaging, the internal defect penetration capability of eddy current detection and the multi-modal data fusion capability of intelligent algorithms. The core value of the application not only lies in the significant improvement of detection performance (such as resolution improvement to 5 μm and classification accuracy ≥ 95%), but also lies in providing a defect detection solution with efficiency, precision and cost advantages for the high-end manufacturing field, which has significant engineering application value and technical foresight. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 It is a functional block diagram of an embodiment of the magnetic material defect detection system based on image recognition of the application.

[0045] Figure 2 It is Figure 1A functional module schematic view of one embodiment of the first obtaining module shown in the first aspect is shown in FIG. 1;

[0046] Figure 3 For Figure 1 A functional module schematic view of one embodiment of the first determining module shown in the first aspect is shown in FIG. 2;

[0047] Figure 4 For Figure 1 A functional module schematic view of one embodiment of the second determining module shown in the first aspect is shown in FIG. 3;

[0048] Figure 5 For Figure 1 A functional module schematic view of one embodiment of the second obtaining module shown in the first aspect is shown in FIG. 4;

[0049] Figure 6 For Figure 1 A functional module schematic view of one embodiment of the third obtaining module shown in the first aspect is shown in FIG. 5.

[0050] BRIEF DESCRIPTION OF THE DRAWINGS

[0051] 10, first obtaining module; 20, first determining module; 30, second determining module; 40, second obtaining module; 50, third obtaining module; 11, first obtaining unit; 12, first determining unit; 13, judging unit; 14, second determining unit; 21, second obtaining unit; 22, third obtaining unit; 23, extracting unit; 24, fourth obtaining unit; 31, fifth obtaining unit; 32, third determining unit; 33, sixth obtaining unit; 34, fourth determining unit; 41, seventh obtaining unit; 42, fifth determining unit; 43, eighth obtaining unit; 44, ninth obtaining unit; 51, tenth obtaining unit; 52, sixth determining unit; 53, eleventh obtaining unit; 54, generating unit. DETAILED DESCRIPTION

[0052] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in combination with the drawings in the specification and specific embodiments.

[0053] As Figure 1As shown, the first embodiment of the present application proposes a magnetic material defect detection system based on image recognition, which comprises a first acquisition module 10, a first determination module 20, a second determination module 30, a second acquisition module 40 and a third acquisition module 50. The first acquisition module 10 is used to control the multi-angle light illumination of the surface of the magnetic material by using a polarized light imaging device, to suppress the specular reflection caused by the strong light reflection characteristics by adjusting the polarizer angle and the light source incident angle, to obtain the original image data with uniform illumination distribution, and to automatically adjust the exposure parameters and the polarization angle if the overexposed area is detected in the image, to obtain the high-quality image suitable for subsequent processing. The first determination module 20 is used to perform multi-scale wavelet transform decomposition according to the obtained high-quality image, to identify the surface defect area by analyzing the texture features and edge information in different frequency components, to determine the defect boundary profile and geometric feature parameters by combining the gradient information if the wavelet coefficient exceeds the preset threshold in the high-frequency component, to judge as the potential defect area, and to determine the depth position and size range of the internal defect by using the signal processing algorithm to analyze the phase difference. The second determination module 30 is used to scan the internal structure of the magnetic material by using an eddy current detection sensor array, to obtain the amplitude and phase change data of the electromagnetic induction signal, to judge that there is an internal defect if the signal amplitude change rate exceeds the reference value, and to determine the depth position and size range of the internal defect by using the signal processing algorithm to analyze the phase difference. The second acquisition module 40 is used to construct a multi-dimensional feature vector according to the geometric feature parameters of the surface defect and the electromagnetic signal characteristics of the internal defect, to train a defect classification model by using a support vector machine algorithm, to determine whether the defect type belongs to a crack, a pore or an inclusion by using the spatial distribution pattern of the feature vector, and to obtain the confidence score of each defect type. The third acquisition module 50 is used to integrate the surface optical detection result and the internal electromagnetic detection result by using a weighted fusion algorithm, to improve the defect confidence of the same area if the surface optical detection and the internal electromagnetic detection both recognize defects in the same area, to determine the final defect detection report by using the spatial position matching and the feature similarity calculation, and to obtain the comprehensive evaluation result containing the defect position, type and severity.

[0054] In the first acquisition module 10, the polarized light imaging device is an optical detection device that uses the polarization characteristics of light to analyze the polarized light reflected or transmitted by the measured object to obtain information about the surface morphology and physical properties of the magnetic material. The multi-angle light control of the surface of the magnetic material using the polarized light imaging device involves using the polarized light imaging device (including polarizing plates, light source systems, and image sensors) to adjust the incident angle of the light source and the deflection angle of the polarizing plate to achieve multi-directional and multi-angle light adjustment on the material surface to suppress the specular reflection caused by the strong light reflection characteristics of the magnetic material and achieve uniform light distribution. In the field of optical detection, the polarization direction of the polarizing plate and the incident angle of the light source are dynamically adjusted to suppress the specular reflection of the surface of the magnetic material, allowing the light to be evenly distributed and obtaining high-quality image data. The core of this technology is to use the polarization characteristics of light and the principles of geometric optics to achieve reflection control through parameter coordination. In this embodiment, the image high-frequency components are decomposed by wavelet transform, potential defect pixels are selected using threshold values, and defect boundary contours are constructed using image gradient information to extract geometric feature parameters. The essence of this technology is to combine signal processing (wavelet analysis) and edge detection (gradient calculation) to achieve accurate positioning and quantitative description of defects.

[0055] In the first determination module 20, the texture features and edge information in different frequency components are extracted by performing multi-scale wavelet transform decomposition on high-quality images, potential defect areas are selected using a pre-set threshold value, and defect boundary contours and geometric feature parameters are determined using gradient analysis. In this embodiment, signal processing and computer vision technology are combined to achieve the entire process from defect detection to quantitative analysis.

[0056] In the second determination module 30, the amplitude and phase change data of the electromagnetic induction signal are collected by scanning the conductive material using a multi-channel eddy current sensor array, and the pre-set reference value and signal processing algorithm are combined to achieve automatic detection, depth positioning, and size quantization of internal defects.

[0057] In the second acquisition module 40, high-dimensional feature vector space is constructed by combining the geometric features of surface defects and the electromagnetic signal features of internal defects, a non-linear classification model is trained using a support vector machine (SVM) algorithm, and the automatic recognition and confidence evaluation of defect types such as cracks, pores, and inclusions are achieved.

[0058] In the third acquisition module 50, the surface optical detection results (such as surface cracks and scratches identified by polarized light imaging) and the internal electromagnetic detection results (such as pores and inclusions found by eddy current detection) are integrated using a weighted fusion algorithm, the reliability of defect detection is improved using spatial position matching and feature similarity calculation, and a comprehensive evaluation report containing position, type, and severity is generated.

[0059] Further, see Figure 2 The image recognition-based magnetic material defect detection system provided by the embodiment further comprises a first acquisition module 10, which comprises a first acquisition unit 11, a first determination unit 12, a judgment unit 13 and a second determination unit 14. The first acquisition unit 11 is configured to perform multi-angle light scanning on the surface of the magnetic material by adjusting the angle of the polarizer and the incident angle of the light source according to the characteristics of polarized light imaging, acquire first group image data from the scanning results, and perform brightness distribution detection on the first group image data to determine whether there is an overexposed area, thereby obtaining a preliminary brightness distribution evaluation result. The first determination unit 12 is configured to, if there is an overexposed area in the first group image data, perform secondary light control on the overexposed area by automatically adjusting the exposure parameter and the angle of the polarizer, acquire second group image data under the adjusted light conditions, and perform brightness uniformity detection on the second group image data to determine whether there is a residual specular reflection area. The judgment unit 13 is configured to, for the residual specular reflection area in the second group image data, perform region segmentation using a preset threshold, extract position information of the reflection interference area from the segmentation result, acquire third group image data by adjusting the local light source intensity and fine-tuning the polarization angle, and determine whether the standard of uniform light distribution is reached. The second determination unit 14 is configured to, according to the uniformity detection result of the third group image data, if there is still a local brightness uneven area, perform pixel-level correction on the local brightness uneven area by using an image processing tool, acquire fourth group image data from the corrected data, and determine whether the requirement for subsequent processing is met.

[0060] The first acquisition unit 11 adjusts the angle of the polarizer and the incident angle of the light source in the detection of the surface of the magnetic material based on polarized light imaging. Polarized light imaging uses the polarization characteristics of light to effectively suppress specular reflection and highlight the micro features of the material surface. Assuming that the initial light source incident angle is set to 45 degrees and the polarizer angle is 0 degrees, the first group image data is acquired by multi-angle light scanning.

[0061] The first determination unit 12 is configured to, when performing brightness distribution detection, if it is found that the brightness value of a certain area in the image exceeds a preset threshold, such as 230 in 255, it is determined to be an overexposed area. This reflects that the preliminary brightness distribution evaluation result is not ideal, which may be due to strong reflection caused by direct light. Specifically, for the overexposed area, secondary light control can be performed by automatically adjusting the exposure parameter, such as shortening the exposure time from 1 / 60 second to 1 / 120 second, and adjusting the polarizer angle from 0 degrees to 30 degrees, to acquire the second group image data.

[0062] If the remaining specular reflection area is still found in the brightness uniformity detection, such as the brightness value of a small area is still higher than the surrounding area by more than 50%, further processing is required. This indicates that the adjustment of the polarization angle and the exposure parameter has not completely eliminated the reflection interference. In the embodiment, the preset threshold is used for region segmentation of the specular reflection area in the second group of image data to extract the reflection interference area position information. Assuming that the threshold is set to a brightness value of 200, after the interference area is segmented, the local light intensity is adjusted, such as reducing the light intensity of the area by 20%, and the polarization angle is fine-tuned to 35 degrees to obtain a third group of image data.

[0063] The calculation formula for region segmentation using the preset threshold is:

[0064] (1)

[0065] In formula (1), represents the region segmentation result at the position coordinate , represents the pixel intensity value at the position, and represents the preset threshold parameter. When the pixel intensity exceeds the threshold, it is marked as a specular reflection area, otherwise it is marked as a non-reflection area. The adjusted light intensity is:

[0066] (2)

[0067] In formula (2), represents the adjusted light intensity, represents the original light intensity, represents the local light source intensity adjustment coefficient, represents the local polarization angle, and represents the polarization angle parameter.

[0068] The light distribution uniformity evaluation index is: (3) In formula (3),

[0069] represents the light distribution uniformity evaluation index, represents the total number of sampling points in the image, represents the light intensity value of the th sampling point, and represents the average value of the light intensity of all sampling points. The closer the light distribution uniformity evaluation index is to 1, the more uniform the light distribution is.

[0070]

[0071]

[0072] ​​​​​​​​​The second determination unit 14 is used for further processing if the uniformity detection shows that the local brightness uneven area still exists, such as the brightness value fluctuation range of a certain area exceeds 10%. For example, for the local brightness uneven area in the third group of image data, pixel-level correction can be performed through an image processing tool. Assuming that the brightness value of a certain area is too high, the fourth group of image data is generated by reducing the pixel value of the area by about 15%. Finally, if the brightness fluctuation range is controlled within 5%, it is considered to meet the requirements of subsequent processing. This hierarchical optimization method not only gradually eliminates overexposure and reflection interference, but also improves image quality and ensures the accuracy of magnetic material surface defect detection. It should be noted that each step of the above technical processing is carried out around the polarization imaging characteristics, and fully utilizes the advantage of polarization light in suppressing reflection. Through multi-angle scanning and parameter adjustment, noise interference in the image can be effectively reduced, and reliable data support is provided for subsequent material surface analysis. Especially in industrial detection, it can significantly improve the defect recognition rate and reduce the risk of misjudgment, and has high practical value.

[0073] Preferably, as shown in Figure 3 The image recognition-based magnetic material defect detection system provided by the embodiment includes a first determination module 20, a second determination module 14, a third determination module 16 and a fourth determination module 18. The first determination module 20 includes a second acquisition unit 21, a third acquisition unit 22, an extraction unit 23 and a fourth acquisition unit 24. The second acquisition unit 21 is used for performing hierarchical processing on high-quality images by using a wavelet transform tool according to the principle of multi-scale decomposition, separating image data of different frequency components from the high-quality images, extracting the details in the high-frequency components, and obtaining initial decomposition data containing texture features and edge information. The third acquisition unit 22 is used for comparing the high-frequency component information in the initial decomposition data with a preset threshold value. If the wavelet coefficient of a certain area exceeds the preset threshold value, it is determined that the area is a first potential defect area, and the specific position distribution data of the first potential defect area is obtained. The extraction unit 23 is used for detecting the changes of the pixels around the area according to the specific position distribution data of the first potential defect area, determining the boundary contour information of the first potential defect area, and extracting the geometric feature data in the boundary contour. The fourth acquisition unit 24 is used for classifying and labeling the potential defect area by using an image processing tool according to the boundary contour information and the geometric feature data, combining the distribution status of the texture features, and obtaining the final defect recognition result.

[0074] The second acquisition unit 21 is used for the application of multi-scale decomposition and wavelet transform. First, multi-scale decomposition is a method of decomposing an image into different frequency components. Wavelet transform, as a commonly used tool, can effectively separate low-frequency and high-frequency information in an image. The low-frequency part usually reflects the overall structure of the image, while the high-frequency part contains detailed information such as texture and edges. In the detection of the surface of magnetic materials, this decomposition method helps to extract the features of small defects from complex images. For example, for layered processing of high-quality images, wavelet transform can be applied to image data decomposition. Assuming that an image has a resolution of 1024x1024 pixels, it can be decomposed into three layers of frequency components by wavelet transform, and the high-frequency components contain the surface fine texture and edge information.

[0075] The third acquisition unit 22 is used for focusing on the distribution of wavelet coefficients in the high-frequency components when extracting the detailed part. Assuming that the peak value of the wavelet coefficients in a certain area is 50, and the preset threshold is 30, this area is preliminarily judged as a potential defect area, so that the position of possible problems can be quickly located.

[0076] The extraction unit 23 is used for further analyzing the pixel changes around the area after determining the potential defect area by using gradient information calculation tools. Assuming that the pixel brightness value around a certain potential defect area changes sharply from 100 to 200, and the gradient value is significantly higher than that of other areas, the boundary profile of the defect can be outlined accordingly. Further extract the geometric feature data within the boundary profile, such as the aspect ratio of the defect area is 2:1, or the area ratio is 0.5% of the total image area, which provides an important basis for subsequent classification.

[0077] The fourth acquisition unit 24 is used for the classification labeling of the defect area, and comprehensive analysis is performed in combination with the texture features and the geometric features. Assuming that the texture features of a certain potential defect area show irregular stripes, and the boundary contour presents an elliptical shape, in combination with historical data comparison, it can be labeled as "crack type defect". In another case, if the texture features are relatively uniform, and the boundary contour is close to a circle, it can be labeled as "bubble type defect". This classification method can provide clear defect type information for subsequent processing. For example, in a specific implementation, the processing for different frequency components can be further optimized. Assuming that some wavelet coefficients in the high frequency component are close to the threshold but do not exceed the threshold, the potential defect area can be re-screened by adjusting the threshold range, such as from 30 to 25. This flexible adjustment helps to avoid missing small defects, while improving the comprehensiveness of the detection. Such a processing method is particularly important in the detection of the surface of magnetic materials, because small defects on the surface of the material often have a great influence on the quality of the final product. For the extraction of the boundary contour information and the analysis of the geometric feature data, an auxiliary tool can also be introduced for verification. Assuming that the boundary contour preliminarily determined through gradient information has a fuzzy area, the boundary can be smoothed by combining image processing tools to ensure the continuity of the contour line.

[0078] This method can improve the accuracy of defect area positioning and lay the foundation for subsequent classification labeling. Such a technical process is of great significance in industrial detection and can effectively support the quality control process.

[0079] Further, referring to Figure 4 The magnetic material defect detection system based on image recognition provided by the embodiment includes a fifth acquisition unit 31, a third determination unit 32, a sixth acquisition unit 33, and a fourth determination unit 34. The fifth acquisition unit 31 is used for comprehensively scanning the internal structure of the material through the sensor array layout, recording the electromagnetic induction signal by using a preset signal acquisition frequency, extracting the amplitude variation rate and phase difference value data, and obtaining the preliminary signal feature distribution. The third determination unit 32 is used for point-by-point comparison of the amplitude variation rate by using a signal processing tool for the preliminary signal feature distribution. If the amplitude variation rate of a certain area exceeds the preset reference value setting, it is determined that there is an internal defect in the area, and the position coordinates of the second potential defect area are determined. The sixth acquisition unit 33 is used for layer-by-layer analysis of the internal defect depth by using the depth positioning parameter in combination with the phase difference value data according to the position coordinates of the second potential defect area, and obtaining the specific depth distribution information of the internal defect. The fourth determination unit 34 is used for boundary division of the defect size range by using a size estimation method for the specific depth distribution information of the internal defect, and determining the complete size range data of the internal defect by comparing the electromagnetic induction signals of adjacent areas through a signal processing tool.

[0080] The fifth acquisition unit 31 is used to discuss the design and application of sensor array layout in the field of material internal structure detection from the comprehensiveness of signal acquisition. The sensor array is usually arranged in a grid or ring shape on the material surface to ensure that every key point of the detection area is covered. By presetting a signal acquisition frequency, such as 1000 times per second to collect electromagnetic induction signals, subtle changes in the material interior can be captured. This high-frequency acquisition method can provide sufficient data support for subsequent signal feature extraction, which is particularly important when detecting small internal defects.

[0081] The third determination unit 32 is used to extract the amplitude change rate and phase difference value by point-by-point analysis of the collected electromagnetic signals through signal processing tools. Assuming that the amplitude change rate of a certain area reaches 1.5 times the reference value, which is significantly higher than the average value of 0.8 times of the surrounding area, it can be preliminarily judged that there is an internal defect in this area, and it is marked as a second potential defect area. This judgment method based on the amplitude change rate can quickly lock the possible problem position and lay the foundation for subsequent in-depth analysis.

[0082] The sixth acquisition unit 33 is used to perform in-depth positioning analysis by combining phase difference value data after determining the position coordinates of the second potential defect area. The phase difference value reflects the delay difference caused by the change of the internal structure of the material during signal propagation. Assuming that the phase difference value of a certain area reaches 30 degrees, while the surrounding area is only 5 degrees, it can be inferred that there is a defect of a certain depth below this area. By layer-by-layer analysis of depth positioning parameters, such as dividing the material thickness into 10 layers and calculating the signal response difference of each layer, the specific depth distribution information of the defect can be obtained, such as the defect being located between the 3rd layer and the 5th layer.

[0083] The fourth determination unit 34 is used to estimate the size of the defect range by comparing the electromagnetic induction signal strength of adjacent areas through signal processing tools. Assuming that the signal strength of the defect area decreases to 60% of the normal area, while the center area is only 30%, the size range of the defect can be roughly outlined, such as 5mm in length and 3mm in width. This size estimation method helps to understand the morphological characteristics of the defect and provides data support for subsequent processing. For example, in actual implementation, the layout density of the sensor array and the selection of signal acquisition frequency need to be adjusted according to the material characteristics. For high-density materials, the number of sensors can be appropriately increased, such as arranging 4 sensors per square centimeter, to improve the signal resolution. This flexible adjustment method can adapt to different detection needs to ensure the accuracy of signal feature distribution and effectively improve the comprehensiveness and reliability of detection.

[0084] Preferably, referring to Figure 5The embodiment provides the magnetic material defect detection system based on image recognition, the second acquisition module 40 includes the seventh acquisition unit 41, the fifth determination unit 42, the eighth acquisition unit 43 and the ninth acquisition unit 44, wherein the seventh acquisition unit 41 is used for carrying out edge detection on the defect area by using an image processing tool according to surface geometric shape data, extracting contour information of the defect from the collected surface image, and obtaining a preliminary geometric feature set;The fifth determination unit 42 is used for combining electromagnetic signal intensity data for the preliminary geometric feature set, extracting the response signal features of the internal defect by using a signal processing tool, constructing a multi-dimensional feature vector containing surface and internal information, and determining a comprehensive feature description result;The eighth acquisition unit 43 is used for carrying out type division on the feature vector by using a classification processing tool if the feature vector in the comprehensive feature description result presents a specific spatial distribution mode, judging that the defect belongs to a crack, a pore or an inclusion, and acquiring classification determination data;The ninth acquisition unit 44 is used for quantitatively processing the confidence of each defect by using a probability calculation tool according to the specific features of the defect type of the classification determination data, and obtaining final score data.

[0085] The seventh acquisition unit 41 is used for carrying out edge detection on the defect area by using an image processing tool for the processing of surface geometric shape data in the field of material surface and internal defect detection, and extracting contour information. Assuming that an irregular area is detected in the surface image of a metal plate, the boundary of the area can be outlined by using an edge detection algorithm, and a geometric feature set with a length-width ratio of 2:1 is obtained preliminarily. This mode is helpful for quickly locking the shape range of the surface defect and laying a foundation for subsequent analysis.

[0086] The fifth determination unit 42 is used for extracting the response signal features of the internal defect by using a signal processing tool for the combination of the preliminary geometric feature set and electromagnetic signal intensity data. Assuming that the electromagnetic signal intensity shows obvious attenuation under the same area of the above metal plate, and is reduced by 40% compared with the surrounding area, it can be preliminarily inferred that there is an internal defect. After a multi-dimensional feature vector containing surface contour and internal signal intensity is constructed, a comprehensive feature description result can be formed. The construction mode of the multi-dimensional feature vector can effectively fuse the surface and internal information, and provide data support for comprehensive analysis of the defect.

[0087] The eighth acquisition unit 43 is used for carrying out type division by using a classification processing tool if the feature vector presents a specific spatial distribution mode in the comprehensive feature description result. Assuming that the signal attenuation mode and surface contour features of a certain area are highly consistent with the typical distribution of a crack in the feature vector analysis, and are greatly different from the mode of pores or inclusions, it can be preliminarily judged that the defect is a crack. This classification mode can quickly distinguish the defect type and provide a clear direction for subsequent processing.

[0088] The ninth acquisition unit 44 is used for quantifying the confidence of the defect type by the probability calculation tool for the classification decision data, and obtaining the final score data. Assuming that in the above crack judgment, the confidence of the crack is 85% by probability calculation, and the confidence of the pores and inclusions is only 10% and 5% respectively, it can be confirmed that the defect is most likely a crack. Such quantification processing mode can improve the reliability of the judgment, avoid subjective misjudgment, and provide data basis for subsequent repair or treatment. For example, for edge detection of surface geometric shape data, the parameter setting of the image processing tool is adjusted according to the roughness of the material surface. Assuming that the surface of the metal plate is relatively rough, the threshold value of edge detection is appropriately increased to filter noise interference and ensure the accuracy of contour extraction. This flexible adjustment mode can adapt to different material characteristics and improve the adaptability of detection. For the construction process of the multi-dimensional feature vector, different signal feature combinations can be selected according to the detection requirements. Assuming that in some high-precision detection scenarios, in addition to the electromagnetic signal strength, the signal phase data can also be added as a feature dimension to further enrich the comprehensive feature description result. This expansion mode can improve the information amount of the feature vector and provide more basis for defect classification. In the classification processing and confidence quantification, historical detection data can be introduced as a reference. Assuming that in the past similar metal plate detection, the feature vector distribution pattern of the crack is highly consistent with the current result, the confidence score can be further improved. This combination of historical data can enhance the credibility of the classification decision and provide more reliable decision support for practical applications.

[0089] Further, see Figure 6The image recognition-based magnetic material defect detection system provided in the embodiment includes a third acquisition module 50, which includes a tenth acquisition unit 51, a sixth determination unit 52, an eleventh acquisition unit 53, and a generation unit 54. The tenth acquisition unit 51 is configured to scan a target region by using an optical imaging tool according to surface detection data, extract preliminary boundary information of a surface defect from a scanned image, perform fine processing on the preliminary boundary information by using an edge detection tool, and obtain accurate feature data of the surface defect. The sixth determination unit 52 is configured to perform signal collection on the target region by using an electromagnetic induction tool according to the requirement of internal detection, extract response features of an internal defect from the signal, and perform region matching on an overlapping region by using a position calibration tool if the response features and the accurate feature data of the surface defect overlap in a spatial position, and determine defect distribution records of the overlapping region. The eleventh acquisition unit 53 is configured to calculate the similarity between surface defect features and internal defect features by using a feature comparison tool according to the defect distribution records, and integrate two detection results by using a weighted fusion tool if the similarity is higher than a preset threshold, and obtain a comprehensive defect feature set. The generation unit 54 is configured to perform type division on a defect by using a classification and discrimination tool according to the comprehensive defect feature set, perform quantitative processing on defect features by using an intensity evaluation tool, determine the specific type and severity of the defect, and generate a defect detection report containing position information.

[0090] The tenth acquisition unit 51 is used in the field of material surface and internal defect detection, and is configured to perform high-resolution scanning on a target region by using an optical imaging tool according to surface detection data, and obtain preliminary image data. It is assumed that a dark area suspected to be a defect is found in the surface scanning of a metal plate, and the boundary of the dark area is roughly outlined by using an image processing software, and a region range with a length-width ratio of 3:1 is obtained. This way provides basic data support for subsequent accurate analysis.

[0091] The sixth determining unit 52 is used for refining the preliminary boundary information. An edge detection tool is used to perform secondary processing on the image to extract more precise contour features. Assuming that in the dark area of ​​the aforementioned metal sheet, by adjusting the sensitivity parameters of the edge detection, noise interference caused by surface roughness is successfully eliminated, and precise data of the defect contour is finally obtained, such as a boundary length of 5 cm and a width of 1.5 cm. This refinement helps to more accurately describe the morphological characteristics of surface defects. In the internal detection stage, an electromagnetic induction tool is used to collect signals from the same area and extract the response features of internal defects. Assuming that below a suspected defect in the metal sheet, the signal intensity is detected to be 30% lower than the surrounding area, and the signal fluctuation exhibits specific frequency characteristics, it is preliminarily inferred that there may be internal voids or cracks. This extraction of signal features lays the foundation for subsequent comparison with surface data. For the spatial overlap judgment of surface and internal features, a position calibration tool is used for region matching. Assuming that the deviation between the precise contour center point coordinates of the surface defect and the center point coordinates of the internal signal anomaly area is only 0.2 cm, it can be confirmed that there is a high degree of overlap between the two, and they are recorded as the same defect distribution area. This matching method helps integrate multi-source data and improves the accuracy of defect location.

[0092] The eleventh acquisition unit 53 is used in the feature comparison stage to calculate the similarity between surface and internal defect features using a feature comparison tool. Assuming the surface contour exhibits elongated strip-like features, and the internal signal attenuation pattern also conforms to the typical characteristics of crack-like defects, with a similarity score of 80%, which is higher than the preset threshold of 60%, it can be preliminarily determined that the two are strongly correlated. This comparison method provides a reliable basis for subsequent fusion.

[0093] The degree of similarity between surface defect features and internal defect features is as follows:

[0094] (4)

[0095] In formula (4), This indicates the degree of similarity between surface defect features and internal defect features. Represents the surface defect feature vector. Represents the feature vector of internal defects. Indicates the number of feature dimensions. Indicates the first Weight coefficients for each feature dimension The angle representing the k-th dimension of the surface defect. Indicates internal defects From the perspective of dimensional features, The function calculates the cosine similarity of two feature vectors across all dimensions.

[0096] The formula for determining the appropriate threshold value for judging whether feature fusion is needed through statistical analysis is:

[0097] (5)

[0098] In formula (5), represents a preset similarity degree determination threshold value, represents the mean value of historical similarity data, represents a threshold adjustment coefficient, represents the standard deviation of historical similarity data.

[0099] The comprehensive defect feature set is:

[0100] (6)

[0101] In formula (6), represents the comprehensive defect feature set, represents the surface detection result feature vector, represents the internal detection result feature vector, represents a weighted fusion coefficient, with a value range of 0 to 1, and formula (6) integrates the two detection results into the final comprehensive feature through linear weighting.

[0102] The generation unit 54 is used for the integration of detection results, combining surface and internal feature data through a weighted fusion tool to form a comprehensive defect feature set. Assuming that the surface feature weight is set to 0.6 and the internal feature weight is 0.4, a multi-dimensional feature description containing shape and signal strength is finally generated. This fusion method can fully reflect the defect characteristics. In defect type classification, a classification discrimination tool is used to analyze the comprehensive feature set. Assuming that the feature data has a higher degree of match with the typical pattern of crack defects, and is significantly different from pores or inclusions, it can be initially classified as a crack type. At the same time, combined with an intensity evaluation tool to quantify the severity of the defect, such as a larger signal attenuation amplitude, it is judged that the depth may be deeper, and a detection report containing position coordinates and severity rating is generated. This combination of classification and quantification provides a clear guide for subsequent processing.

[0103] The defect detection report containing position coordinates is:

[0104] (7)

[0105] In formula (7), represents a defect detection report with position coordinates represents the identified defect type code, represents the quantified severity value, ​the precise coordinate position of the defect center point, the area range of defect influence, the detection timestamp.

[0106] The fusion defect confidence is calculated by the following formula:

[0107] (8)

[0108] In formula (8), denotes the fusion defect confidence at position denotes the defect confidence of surface optical detection at position denotes the defect confidence of internal electromagnetic detection at position denotes the defect confidence of internal electromagnetic detection at position denotes the weight coefficient of optical detection, denotes the weight coefficient of electromagnetic detection, denotes the confidence boost factor, denotes the indication function when both detection modalities identify a defect at the same position. The spatial position matching similarity is calculated by the following formula:

[0109]

[0110] (9)

[0111] In formula (9), denotes the spatial position matching similarity between defect and defect , and denote the spatial coordinates of defect , and denote the spatial coordinates of defect , denotes the standard deviation parameter of spatial position matching, used to control the sensitivity of spatial matching.

[0112] The final defect detection report is:

[0113] (10)

[0114] In formula (10), denotes the final defect detection report, denotes the defect position information, denotes the defect type, denotes the defect severity, denotes the confidence of the th candidate defect, denotes the​​ a feature similarity score of the candidate defect, a multimodal matching score of the candidate defect, a multimodal matching score of the candidate defect, a weight coefficient of different evaluation indexes.

[0115] Compared with the prior art, the magnetic material defect detection system based on image recognition provided by the embodiment has the following beneficial effects:

[0116] I. Synergistic effect of multimodal detection technology

[0117] 1. Breakthrough of surface defect detection capability

[0118] Strong light suppression and uniform imaging: through dynamic adjustment of the angle of the polarizer and the incident angle of the light source, the specular reflection on the surface of the magnetic material (such as metal alloy) is effectively suppressed, solving the problem of defect omission caused by overexposure in traditional optical detection. For example, in the detection of high-reflectivity silicon steel sheets, the recognition rate of surface scratches and micro-cracks can be improved from 70% in traditional visual detection to more than 95%.

[0119] Multi-scale feature precise extraction: after wavelet transform decomposes the image into multiple frequencies, the high-frequency component can sensitively capture surface texture mutations (such as high-frequency noise of crack edges), and combined with gradient analysis, it can accurately outline the defect profile (positioning error ≤0.1 pixels), which improves the recognition adaptability of different scale defects (from micron-level scratches to millimeter-level cracks) by 80% compared with traditional edge detection algorithms (such as Canny).

[0120] 2. Depth and precision improvement of internal defect detection

[0121] Three-dimensional scanning capability of eddy current array: the sensor array determines the existence of internal defects through the rate of change of electromagnetic signal amplitude (such as > 5% of the reference value), and realizes defect depth positioning (error ≤1mm) and size measurement (such as pore diameter error ≤5%) through phase difference analysis (such as phase difference >10°), breaking through the bottleneck of limited detection depth (traditional single probe ≤5mm) of single eddy current sensor, and the detection depth can reach more than 20mm.

[0122] Anti-interference ability of signal processing: through filtering algorithm to eliminate environmental noise (such as power frequency interference), the signal-to-noise ratio of electromagnetic signal is improved to more than 20dB, significantly reducing the misjudgment rate caused by signal noise points (traditional method misjudgment rate ≥10%, the present invention ≤3%).

[0123] II. Defect analysis upgrade driven by intelligent algorithm

[0124] 1. Multi-dimensional feature fusion and classification accuracy ​​

[0125] Cross-modal feature vector construction: Geometric parameters of surface defects (such as length, curvature, and roughness) and electromagnetic characteristics of internal defects (such as amplitude change rate and phase gradient) are integrated into feature vectors of more than 10 dimensions. High-precision classification of defect types is achieved using a nonlinear kernel function (such as the RBF kernel) of a Support Vector Machine (SVM). Experimental data show that the classification accuracy for cracks, pores, and inclusions reaches 98%, 92%, and 95%, respectively, representing an improvement of more than 25% compared to the traditional thresholding method.

[0126] Quantitative evaluation of confidence scores: Calculate the confidence score of each defect type (e.g., confidence score of crack ≥90%) by the distribution density of feature vectors in space (e.g., intra-class distance <0.5), to provide data support for subsequent decision-making and avoid the ambiguity of human experience judgment.

[0127] 2. Decision reliability of weighted fusion

[0128] Dual verification of space and features: When both surface optical and internal electromagnetic detection identify defects in the same area (spatial coordinate deviation ≤ 0.5 mm), the overall confidence level of the area is improved by weighting factors (e.g., surface detection weight 0.6, electromagnetic detection weight 0.4), effectively reducing the risk of cross-modal missed detection.

[0129] Comprehensive assessment of defect severity: Combining the geometric dimensions of surface defects (e.g., crack length > 2 mm) and the depth of internal defects (e.g., ≥ 5 mm), a severity score (1-10 points) is calculated using the Analytic Hierarchy Process (AHP) to provide a quantitative basis for defect repair or scrapping decisions.

[0130] III. Improvement of Detection Efficiency and Industrial Applicability

[0131] 1. Automated processes and online testing capabilities

[0132] Real-time feedback and parameter self-adjustment: The system can adapt to different reflective conditions without manual intervention by automatically identifying overexposed areas of the image (response time ≤ 50ms) and dynamically adjusting the polarization angle and exposure parameters. The detection efficiency is increased to more than 50 pieces / minute, which is 10 times higher than the traditional manual adjustment mode.

[0133] Multi-station parallel inspection deployment: Polarized light imaging and eddy current array scanning can be performed simultaneously, and with the help of the production line conveyor system, the entire process of "feeding-inspection-sorting" can be automated, which is suitable for online quality control of large batches of magnetic materials (such as steel plates and pipes).

[0134] 2. Adaptability and cost optimization in complex scenarios

[0135] Multi-material compatible detection: By switching the type of polarized light source (such as linearly polarized light / circularly polarized light) and the frequency of eddy current sensor (such as 10kHz-1MHz), it can meet the detection needs of ferromagnetic materials (such as carbon steel) and non-ferromagnetic conductive materials (such as aluminum alloy), reducing the cost of repeated investment in enterprise equipment.

[0136] Defect tracing and process optimization: The correlation analysis of defect location, type and processing parameters (such as rolling temperature, cutting speed) recorded by the system can help the process department quickly locate the root cause of the problem (such as finding that the internal porosity of a batch of materials is related to the fast cooling rate of casting), and shorten the process debugging cycle by more than 30%.

[0137] Four, the industry demonstration value of technological innovation

[0138] 1. Integrated innovation of cross-disciplinary detection technology

[0139] Deeply integrate polarized optics, electromagnetic induction and machine learning algorithms to form a complete technology chain of "physical field perception-feature extraction-intelligent decision-making", providing a new paradigm for metal material defect detection with cross-modal collaboration, which can be reused in fields such as aerospace, nuclear power, and automobile manufacturing that require extremely high detection accuracy.

[0140] 2. Foundation construction of intelligent quality control

[0141] The detection results are uploaded to the MES system in real time through the industrial bus (such as Profinet), and a defect prediction model (such as LSTM-based defect rate trend analysis) is established based on historical data, promoting the transformation of quality control from "post-detection" to "prevention", and helping enterprises build a digital and intelligent quality management system.

[0142] In summary, the magnetic material defect detection system based on image recognition provided in this embodiment realizes the leap from "single modal local detection" to "full-dimensional intelligent diagnosis" of magnetic material defect detection through the triple superposition of the surface detail capture capability of polarized light imaging, the internal defect penetration capability of eddy current detection, and the multi-modal data fusion capability of intelligent algorithms. Its core value not only lies in the significant improvement of detection performance (such as resolution improved to 5μm, classification accuracy ≥95%), but also in providing a defect detection solution with efficiency, precision and cost advantage for high-end manufacturing fields, which has significant engineering application value and technical foresight.

[0143] While the preferred embodiments of the application have been described, additional variations and modifications can be made to these embodiments by those skilled in the art once they have the benefit of the foregoing description without departing from the spirit and scope of the application. Accordingly, it is intended that the appended claims be interpreted as including all such variations and modifications as fall within the spirit and scope of the application. It is further intended that the disclosure of all such modifications and variations be included within the scope of the application, the terms used herein being defined solely for purposes of the description being applied thereto unless otherwise indicated.

Claims

1. An image recognition based magnetic material defect detection system, characterized by, The method comprises the following steps: A first acquisition module (10) is configured to control multi-angle light irradiation on the surface of the magnetic material by using a polarized light imaging device, suppress specular reflection caused by strong light reflection characteristics by adjusting the angle of a polarizer and the incident angle of a light source, acquire original image data with uniform light distribution, and automatically adjust exposure parameters and the angle of the polarizer if overexposed areas are detected in the image to obtain high-quality images suitable for subsequent processing; A first determination module (20) is configured to perform multi-scale wavelet transform decomposition based on the acquired high-quality images, identify surface defect areas by analyzing texture features and edge information in different frequency components, and determine defect boundary profiles and geometric feature parameters by combining gradient information if wavelet coefficients exceed a preset threshold in high-frequency components; A second determination module (30) is configured to scan the internal structure of the magnetic material by using an eddy current detection sensor array, acquire amplitude and phase change data of electromagnetic induction signals, determine the presence of internal defects if the signal amplitude change rate exceeds a reference value, and analyze phase differences by using a signal processing algorithm to determine the depth position and size range of internal defects; A second acquisition module (40) is configured to construct a multi-dimensional feature vector based on the geometric feature parameters of surface defects and the electromagnetic signal features of internal defects, train a defect classification model by using a support vector machine algorithm, determine whether the defect type belongs to a crack, a pore, or an inclusion by using the spatial distribution pattern of the feature vector, and obtain a confidence score for each defect type; A third acquisition module (50) is configured to integrate surface optical detection results and internal electromagnetic detection results by using a weighted fusion algorithm, improve the defect confidence of a region if defects are identified in the region by both surface optical detection and internal electromagnetic detection, determine a final defect detection report by using spatial position matching and feature similarity calculation, and obtain a comprehensive evaluation result containing defect position, type, and severity.

2. The image recognition based magnetic material defect detection system of claim 1, wherein, The first acquisition module (10) comprises: A first acquisition unit (11) is configured to perform multi-angle light scanning on the surface of the magnetic material by adjusting the angle of a polarizer and the incident angle of a light source based on the characteristics of polarized light imaging, acquire a first set of image data from the scanning results, and perform brightness distribution detection on the first set of image data to determine whether there are overexposed areas and obtain a preliminary brightness distribution evaluation result; A first determination unit (12) is configured to perform secondary light control on the overexposed areas by automatically adjusting exposure parameters and the angle of the polarizer if there are overexposed areas in the first set of image data, acquire a second set of image data under the adjusted light conditions, and perform brightness uniformity detection on the second set of image data to determine whether there are residual specular reflection areas; A judgment unit (13) is configured to perform region segmentation on the residual specular reflection areas in the second set of image data by using a preset threshold, extract position information of the reflection interference areas from the segmentation results, acquire a third set of image data by adjusting local light intensity and fine-tuning the angle of the polarizer, and determine whether the standard of uniform light distribution is met. The second determining unit (14) is configured to, if there still exists a local brightness uneven area according to the uniformity detection result of the third group of image data, perform pixel-level correction on the local brightness uneven area by using an image processing tool, obtain a fourth group of image data from the corrected data, and determine whether the fourth group of image data meets the requirement of subsequent processing.

3. The image recognition based magnetic material defect detection system of claim 2, wherein, In the judging unit (13), a preset threshold is used in a calculation formula for region segmentation, and the calculation formula is as follows: wherein, represents a position coordinate a region segmentation result at the position, represents a pixel intensity value at the position, represents a preset threshold parameter; when the pixel intensity exceeds the threshold, it is marked as a specular reflection region, otherwise it is marked as a non-reflection region; The adjusted light intensity is: wherein, represents the adjusted light intensity, represents the original light intensity, represents the local light source intensity adjustment factor, represents the local polarization angle, represents the polarizer angle parameter; The light distribution uniformity evaluation index is: wherein, denotes the illumination distribution uniformity evaluation index, denotes the total number of sampling points in the image, denotes the illumination intensity value of the th sampling point, denotes the average value of the illumination intensity of all sampling points.

4. The image recognition based magnetic material defect detection system of claim 1, wherein, The first determining module (20) comprises: The second obtaining unit (21) is configured to, according to the principle of multi-scale decomposition, separate image data of different frequency components from the high-quality image by using a wavelet transform tool, extract a detail part in a high-frequency component, and obtain initial decomposition data containing texture features and edge information. The third obtaining unit (22) is configured to, for the high-frequency component information in the initial decomposition data, compare the high-frequency component information with a preset threshold, and if a wavelet coefficient of a region exceeds the preset threshold, determine that the region is a first potential defect region, and obtain specific position distribution data of the first potential defect region. The extracting unit (23) is configured to, according to the specific position distribution data of the first potential defect region, detect a change of pixels around the region by using a gradient information calculation tool, determine boundary contour information of the first potential defect region, and extract geometric feature data in the boundary contour. The fourth obtaining unit (24) is configured to, by using the boundary contour information and the geometric feature data, and in combination with a distribution state of the texture features, classify and label the potential defect region by using an image processing tool, and obtain a final defect recognition result.

5. The image recognition based magnetic material defect detection system of claim 4, wherein, The second obtaining unit (21) is specifically configured to separate image data of a low-frequency component and image data of a high-frequency component from the high-quality image by using a wavelet transform tool.

6. The image recognition based magnetic material defect detection system of claim 1, wherein, The second determining module (30) comprises: The fifth obtaining unit (31) is configured to comprehensively scan an internal structure of a material by using a sensor array layout, record the electromagnetic induction signal by using a preset signal acquisition frequency, extract amplitude variation rate and phase difference value data from the electromagnetic induction signal, and obtain initial signal feature distribution. The third determining unit (32) is configured to, for the initial signal feature distribution, perform point-by-point comparison on the amplitude variation rate by using a signal processing tool, if an amplitude variation rate of a region exceeds a preset reference value, determine that the region has an internal defect, and determine position coordinates of a second potential defect region. The sixth obtaining unit (33) is configured to, according to the position coordinates of the second potential defect region, in combination with the phase difference value data, perform layer-by-layer analysis on a depth of the internal defect by using a depth positioning parameter, and obtain specific depth distribution information of the internal defect. The fourth determining unit (34) is configured to, for the specific depth distribution information of the internal defect, perform boundary division on a defect size range by using a size estimation method, compare electromagnetic induction signals of adjacent regions by using a signal processing tool, and determine complete size range data of the internal defect.

7. The image recognition based magnetic material defect detection system of claim 1, wherein, The second acquisition module (40) comprises: A seventh acquisition unit (41) configured to perform edge detection on the defect area by using an image processing tool according to the surface geometry data, extract contour information of the defect from the collected surface image, and obtain a preliminary geometry feature set; A fifth determination unit (42) configured to perform feature extraction on a response signal of the internal defect by using a signal processing tool in combination with the electromagnetic signal intensity data for the preliminary geometry feature set, construct a multi-dimensional feature vector containing surface and internal information, and determine a comprehensive feature description result; An eighth acquisition unit (43) configured to perform type division on the feature vector in the comprehensive feature description result by using a classification processing tool if the feature vector presents a specific spatial distribution mode, judge whether the defect belongs to a crack, a pore or an inclusion, and acquire classification determination data; A ninth acquisition unit (44) configured to quantitatively process a confidence degree of each defect by using a probability calculation tool according to the classification determination data and specific features of the defect type, and obtain final score data.

8. The image recognition based magnetic material defect detection system of claim 1, wherein, The third acquisition module (50) comprises: A tenth acquisition unit (51) configured to perform scanning on the target area by using an optical imaging tool according to the data of surface detection, extract preliminary boundary information of the surface defect from the scanned image, perform refinement processing on the preliminary boundary information by using an edge detection tool, and obtain accurate feature data of the surface defect; A sixth determination unit (52) configured to perform signal collection on the target area by using an electromagnetic induction tool for internal detection requirements, extract a response feature of the internal defect from the signal, and perform region matching by using a position calibration tool if the response feature and the accurate feature data of the surface defect overlap in spatial position, and determine a defect distribution record of the overlapping region; An eleventh acquisition unit (53) configured to calculate a similarity degree of the surface defect feature and the internal defect feature by using a feature comparison tool according to the defect distribution record, and integrate two detection results by using a weighted fusion tool if the similarity degree is higher than a preset threshold, and obtain a comprehensive defect feature set; A generation unit (54) configured to perform type division on the defect by using a classification discrimination tool for the comprehensive defect feature set, quantitatively process the defect feature by using an intensity evaluation tool, judge a specific type and severity of the defect, and generate a defect detection report containing position information.

9. The image recognition based magnetic material defect detection system of claim 8, wherein, In the sixth determination unit (52), the electromagnetic induction tool comprises an eddy current detection sensor array.

10. The image recognition based magnetic material defect detection system of claim 8, wherein, In the eleventh acquisition unit (53), the similarity degree of the surface defect feature and the internal defect feature is: wherein, represents the similarity degree of the surface defect feature and the internal defect feature, represents the surface defect feature vector, represents the internal defect feature vector, represents the number of feature dimensions, represents the weight coefficient of the th feature dimension, represents the angle of the kth dimension of the surface defect, represents the angle of the th dimension of the internal defect, the function calculates the cosine similarity of the two feature vectors in each dimension. A formula for determining a suitable threshold value for judging whether feature fusion is needed is: wherein, represents a preset similarity degree determination threshold value, represents a mean value of the historical similarity data, represents a threshold adjustment coefficient, represents a standard deviation of the historical similarity data; The comprehensive defect feature set is: wherein, represents a comprehensive defect feature set, represents a surface detection result feature vector, represents an internal detection result feature vector, represents a weighted fusion coefficient.

Images