Comprehensive detection system and method for steel plate defects

Abstract

The invention provides a comprehensive detection system and method for steel plate defects, and relates to the technical field of steel plate defect detection, the comprehensive detection system comprises an infrared sensor, a synchronous laser encoder, a Hall sensor, a control device, an image acquisition device, a semi-shielding excitation device and a server, and the infrared sensor is connected with the control device. According to the comprehensive detection system and method for the steel plate defects, key subsystems such as image acquisition, magnetic flux leakage detection and magnetic memory detection are organically integrated through modular and complete-set design, and a set of detection device which is high in industrialization degree and flexible to deploy is formed under the assistance of a multi-stage filtering noise reduction and intelligent image processing algorithm. The device not only can stably operate under complex working conditions, but also can quickly position and classify internal and external defects of the steel plate, so that the detection accuracy is remarkably improved, and the risks of missing detection and error detection are effectively reduced, thereby meeting the strict requirements of an efficient production line on quality control.

Description

Technical Field

[0001] This invention relates to the field of steel plate defect detection technology, specifically to a comprehensive detection system and method for steel plate defects. Background Technology

[0002] In various engineering constructions, long-span steel bridges, sheet piles, barrel bracing structures, and steel pipe structures are common. These types of structures are used for many years, and under long-term cyclic stress loads and natural and physical erosion, various defects inevitably appear. Due to harsh inspection conditions, many defects are difficult to detect manually, resulting in inadequate maintenance measures, which in turn affect the structure's performance and lifespan.

[0003] Currently, steel plate testing methods such as penetrant testing, magnetic particle testing, and ultrasonic flaw detectors are relatively complex to operate in actual engineering, have low testing efficiency, and are unable to detect defects on the inner surface that cannot be directly observed. Traditional magnetic memory testing is easily affected by environmental magnetic fields, while magnetic flux leakage testing is difficult to analyze stress state, and neither of these methods can determine whether the defect is located on the inner or outer surface of the steel plate in engineering applications.

[0004] Therefore, this invention extracts various magnetic characteristic parameters by double magnetizing the steel plate, uses the wavelet mode maxima method and the second derivative zero-crossing method to evaluate stress, and corrects the defect edges twice to improve detection accuracy and effectively prevent stress damage; introduces an image acquisition device (12) to determine surface defects, which effectively reduces the missed detection of internal defects; and uses a combination of image detection and magnetic flux leakage detection to improve detection speed and accuracy, providing a new method for comprehensive detection of steel plate defects. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a comprehensive detection system and method for steel plate defects. The technical problem this invention aims to solve is: how to achieve a high-precision, low-false-detection, and low-missing-detection comprehensive safety detection system and method for steel plate defects by using multiple filtering and noise reduction, multiple precise edge judgments, and combined with overall stress state visualization analysis.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a comprehensive detection system for defects in steel plates, comprising:

[0007] Infrared sensors, synchronous laser encoders, Hall sensors, control devices, image acquisition devices, semi-shielded excitation devices, and servers.

[0008] The infrared sensor is connected to the control device;

[0009] The synchronous laser encoder is connected to the control device, and the encoder transmits the synchronous trigger time signal to the control device.

[0010] The Hall sensor is connected to the controller. When the Hall sensor receives the acquisition signal from the controller, it starts to acquire the magnetic field information of the current position and transmits the magnetic field information to the control device.

[0011] The image acquisition device and the control device are connected. When the image acquisition device receives the acquisition signal from the control device, it starts the camera to acquire image information of the steel plate surface and sends it to the control device.

[0012] The control device is connected to an infrared sensor, a synchronous laser encoder, a Hall sensor, an image acquisition device, and a server, respectively.

[0013] The semi-shielded active excitation device uses neodymium iron boron permanent magnets to saturate the detection area of ​​the steel plate, and the semi-shielded shell reduces the interference of the permanent magnets on the leakage magnetic field at the rear.

[0014] The server receives the associated data uploaded by the control device and stores the associated data in three databases.

[0015] Preferably, the neodymium iron boron permanent magnet is 1.8T, and the integrated testing system also includes an outer frame support.

[0016] Preferably, the laser emitter is fixed on the outer frame support and is used to project laser stripes of a specific frequency onto the upper surface of the steel plate. The illumination angle can be adjusted, and the height can also be adjusted by relying on the slide rails at both ends. The 3D industrial camera is deployed on the outer frame support.

[0017] Preferably, the semi-shielded excitation device consists of a semi-shielded magnetic shielding barrel and an internal neodymium iron boron permanent magnet.

[0018] A comprehensive detection method for defects in steel plates includes the following steps:

[0019] S1: Turn on the external power supply to power the entire equipment;

[0020] S2: The inspection trolley moves at a constant speed along the preset inspection line on the steel plate to be inspected. When the infrared sensor detects that it has reached the area to be inspected, it sends an arrival signal back to the control device.

[0021] S3: The control device starts the synchronous laser encoder based on the preset parameters and the system signal that it has entered the preset detection area, so that it outputs synchronous trigger signals at the set time intervals.

[0022] S4: When the synchronous trigger signal is issued, the control device simultaneously triggers the front and rear Hall sensors to collect magnetic field data, the image acquisition device collects image data of the steel plate surface, and transmits the collected information to the controller in real time.

[0023] S5: The controller associates and packages time information, magnetic field data, and image data, and transmits the data collected at each time point to the server wirelessly.

[0024] S6: Data acquisition ends when the rear infrared sensor detects the edge of the detection area;

[0025] S7: The server adjusts the actual location of the packaged and transmitted information;

[0026] S8: The server performs noise reduction processing on the magnetic field data collected by the front-end Hall sensor, then performs defect edge positioning and non-defect interval positioning, and then proceeds to step S9.

[0027] S9: The server selects the image processing of the defect area to determine whether there are surface defects. After all sub-regions are determined, proceed to step S10.

[0028] S10: The server expands the sub-intervals of surface defects, and then extracts the magnetic field data corresponding to the expanded location intervals to locate new defects.

[0029] S11: The server stores all the information of each sub-interval, maps the edge coordinates of different intervals to the axial detection line, and marks the interval boundaries with line segments of different colors.

[0030] S12: The server sends the steel plate defect location detection results to the on-site large screen via wired communication. After receiving the detection information, the on-site large screen displays it on-site, and on-site employees can observe the stress change and defect location information of the detected steel plate through the on-site large screen.

[0031] S13: Employees log into the system using handheld mobile devices such as PADs. Based on the quantitative analysis information of steel plate defects stored on the host server, they can view the detailed situation of defect plate detection. According to the stress state distribution of the steel plate and the location of the defects, they can arrange personnel to carry out corresponding grinding, welding and other treatments.

[0032] Preferably, step S8 includes the following steps:

[0033] Step S8.1: The unsaturated magnetized magnetic field strength data acquired by the front Hall sensor along the one-dimensional coordinate x is defined as a discrete sequence:

[0034] B = {B(x1), B(x2), ..., B(x...} N )}

[0035] Where x i =dx, where dx is the sampling point interval in mm, N is the total number of sampling points, and the detection range is x∈[0,L], L=N·dx;

[0036] For the defined discrete sequence, calculate the mean:

[0037]

[0038] Then the data is zero-mean normalized: B'(x) i )=B(x i )-μ, resulting in a new sequence B'={B'(x1),B'(x2),…,B'(x... N )};

[0039] 8.2: Using the Mexican Hat Wavelet: Perform J-level discrete wavelet decomposition on B'(x) and use the Mallat algorithm to obtain the low-frequency approximate component A. J and high-frequency detail components D1, D2, ..., D J , corresponding to scale a j =2 j dx(j=1,2...J, e.g. J=5 corresponds to scales 1,2,4,8,16dx);

[0040] Then, the standard deviation of the noise is estimated using the first layer of high-frequency components:

[0041]

[0042] Where D1: the first layer of high-frequency detail components, median(|D1|): the median of the absolute values ​​of the elements in D1, and 0.6745: the conversion coefficient between the median and standard deviation of the Gaussian distribution;

[0043] For the high-frequency detail component D of each layer j Using unbiased likelihood estimation, the threshold λ for wavelet-based denoising is calculated:

[0044]

[0045] M = len(D) j ):D j The length of the high-frequency component in the j-th layer, d i :D j The i-th coefficient, sgn(d) i ): Sign function, 1 when di>0, 0 when di=0, -1 when di<0; Card counting function, counts the number of coefficients whose absolute value is greater than the threshold λ, iterates through λ≧0 and takes the λ that minimizes R(λ) as the threshold λ. SURE,j ;

[0046] Soft thresholding is applied to the high-frequency component Dj in each layer. The formula for high-frequency component shrinkage is as follows:

[0047]

[0048] The wavelet is reconstructed layer by layer, and the formula is:

[0049]

[0050] S8.3: Calculate the wavelet coefficient modulus at various scales:

[0051]

[0052] a j =2 j ·dx: The scale corresponding to the j-th level decomposition, The Mexican hat wavelet after translation and scaling, and the signal in x i Convolution is performed at a specific location to extract mutation information at that scale.

[0053] For position x i If M(a) j ,x i )≥M(a j ,x i-1 And M(a) j ,x i )≥M(a j ,x i+1 If x i This represents the maximum modulus point at this scale, corresponding to the location of a signal abrupt change.

[0054] From the largest scale a J Begin by recording the coordinates of all modulus maxima at this scale. For each At the sub-large scale a J-1 Search If a point with a modulus maximum exists within the range, retain that point; repeat the above process until the smallest scale a1 is reached, ultimately retaining points that appear in ≥2 consecutive scales, i.e., points that simultaneously exist in scale a1. j and a j+1 scale;

[0055] S8.4: Sort the selected candidate points according to their x-coordinates to form a sequence {X = x1, x2, ..., x_M}; merge points whose adjacent distance is less than or equal to 2dx to form clusters C1, C2, ..., C_M. k The mean of each cluster is used as the final edge coordinates:

[0056]

[0057] Specifically, the determination of the left and right edges of the sub-defects involves: the left edge corresponding to a sudden change in the magnetic field signal from low to high; and the right edge corresponding to a sudden change in the magnetic field signal from high to low, which serves as the coordinate system for calibrating the edge of the defect region [x]. left x right ];

[0058] S8.5: If K defects are detected, there are 2K edge coordinates x1. <x2<…<x 2k Then the non-defect area is:

[0059]

[0060] The non-defect region is divided into S according to the length ΔL. m =[x m ,x m +ΔL],x m This marks the starting point of the non-defect area.

[0061] For non-defect sub-intervals with intervals, magnetic characteristic parameters are selected, including:

[0062] Wavelet energy ratio:

[0063]

[0064] Among them, E(a) j ): Scale a j Wavelet energy in the non-defect region;

[0065] Signal variance:

[0066]

[0067] Where, N m :S m number of interior points;

[0068] Modulus maximum density: K m For S m Number of internal modulus maxima, L m =ΔL, unit: mm;

[0069] For wavelet energy ratio R m Setting a threshold T under different stress states R1 ,T R2 ,T R3 Similarly, for the other two parameters, T is set... σ1 T σ2 ,T σ3 ,T D1 ,T D2 ,T D3 The threshold value was determined through standard specimen experiments.

[0070] Elastic phase: R m <T R1 , D m <T D1 ;

[0071] Yield stage: T R1 ≤R m <T R2 , T D1 ≤D m <T D2 ;

[0072] Enhancement Phase: T R2 ≤R m <T R3 , T D2 ≤ <T D3 ;

[0073] Neck constriction phase: R m >T R3 , D m >T D3 .

[0074] Preferably, S9 specifically includes the following steps:

[0075] 9.1: For the edge coordinates of the calibrated sub-defect region [x] left x right ], calculate the center position of the interval The center pixel coordinates of the image at the center position are (x center_px ,y center_px ), with x center_px Using the center as the reference point, select a rectangular region of size W*H as the Region of Interest (ROI), denoted as I. ROI ;

[0076] S9.2 performs a single-channel grayscale conversion on the captured image. The formula for RGB to grayscale conversion is:

[0077] I(x,y)=0.299·R(x,y)+0.587·G(x,y)+0.114·B(x,y)

[0078] R(x,y), G(x,y), B(x,y) are the three-channel pixel values ​​of the color image at position (x,y), and I(x,y) are the converted grayscale values.

[0079] 9.3: Perform Gaussian filtering on the output grayscale image for noise reduction. Two-dimensional Gaussian distribution function:

[0080]

[0081] σ is the standard deviation, typical value: (σ=2.0) The kernel size takes an odd number (e.g. 5×5), that is, x,y∈{-2,-1,0,1,2};

[0082] For each pixel (i,j) in the image, calculate its weighted average with the Gaussian kernel:

[0083]

[0084] 9.4: For the ROI region after noise reduction, divide it into 8*8 sub-regions, count the gray values ​​(0~255) of 8×8=64 pixels one by one, record the number of times each gray level appears, and output an array h[i] of length 256, where h[i] is the number of pixels corresponding to the gray level and i is the gray level;

[0085] For the output array, set the clipLimit value to limit the histogram height, and the proportion of pixels of each gray level to the total number of pixels in the sub-region shall not exceed clipLimit / 255;

[0086] clipLimit: The threshold for probability density, measured in pixels per gray level. It indicates the maximum number of pixels per gray level. If the number of pixels for a gray level in a sub-region exceeds clipLimit, the excess pixels are evenly distributed across all 256 gray levels. clipLimit is typically set to 1.0–3.0.

[0087] For all gray levels, the formula for calculating the cumulative distribution function is:

[0088]

[0089] clip(h(i),clipLimit): Limits the number of pixels for each gray level to no more than clipLimit, i.e., h(i) after clipLimit processing;

[0090] Perform grayscale mapping on each sub-region, assigning a new grayscale value to the pixels within that sub-region. The calculation formula is as follows:

[0091]

[0092] I blur (x,y): The grayscale value of this pixel after Gaussian filtering; CDF(I blur (x,y): The number of accumulated pixels for this gray level;

[0093] Image I obtained after CLAHE processing clahe This is a single-channel grayscale image, where each pixel value represents the grayscale intensity at that location.

[0094] S9.5: Horizontal template G using the Sobel operator x and vertical template G y Used to detect edges in the horizontal and vertical directions respectively.

[0095]

[0096] For image I clahe For each pixel (i,j) in G, take its surrounding 3×3 neighborhood pixel block, and combine this neighborhood pixel block with G. x The elements at corresponding positions in the template are multiplied together, and then all products are summed to obtain the horizontal gradient value G of that pixel. x Repeat the above operation for each pixel (i,j) in the image to obtain the horizontal gradient image G. x Similarly, the vertical gradient image G is obtained. y ;

[0097] For each pixel (i,j), calculate its gradient magnitude using the formula:

[0098]

[0099] Repeat the above operation for each pixel in the image to obtain the gradient magnitude image G;

[0100] For each pixel (i,j), calculate its gradient direction using the formula:

[0101]

[0102] The gradient direction is quantized to four main directions: 0°, 45°, 90°, and 135°.

[0103] Based on the set high threshold T of the gradient magnitude h =150 and low threshold T l =50, for each pixel (i,j) in image G', if G'(i,j)≥T h If T is an edge pixel, then the pixel is marked as a strong edge pixel; l ≤G'(i,j) <T h If T l If G'(i,j) is greater than the edge pixel, then the pixel is marked as a non-edge pixel.

[0104] Iterate through all pixels (i,j) marked as weak edges, and check if there are any strong edge pixels in the 8-neighborhood of each pixel. If there are strong edge pixels, mark the weak edge pixel as an edge pixel; otherwise, mark it as a non-edge pixel.

[0105] 9.6: For the output binary edge map, use OpenCV's findContours function to extract all closed contours, and map each contour to a coordinate list {(x1,y1),(x2,y2),…,(x... n ,y n)};

[0106] For a closed contour to be defined, x must satisfy... n+1 =x1,y n+1 =y1, formula for contour area:

[0107]

[0108] Based on the actual minimum detectable size of the defect, a minimum area threshold A is set. threshold Values ​​less than this are considered noise.

[0109] 9.7: Convert the grayscale ROI image to RGB format, use OpenCV's drawContours function to draw the contours, select red as the color (RGB value (255,0,0), set the line width to d=2 pixels, set the contour index to -1, draw all contours that meet the conditions, output a colored ROI image with red contour markers for surface defects, and add the label "Surface Defect", output the original grayscale ROI image for inner surface defects, and add the label "Inner Surface Defect".

[0110] Preferably, step S10 specifically includes the following steps:

[0111] S10.1 For image recognition, the defect coordinates [x] are determined to be surface defects. left ,x right Expanding the range at both ends by a percentage p% yields the target detection interval:

[0112]

[0113] Extract the rear Hall sensor (2) in [x left x right The magnetic field data sequence within the interval is B(x) = [B1, B2, ..., B...]. n ], where B i For position x i Magnetic flux density at the location;

[0114] 10.2: The magnetic field data is smoothed using a one-dimensional Gaussian kernel G(k,σ). The kernel function is:

[0115]

[0116] With a kernel size of 5 and a standard deviation σ = 1.0, the filtered data is as follows:

[0117]

[0118] The filtered data is mapped to the [0,1] interval to eliminate amplitude differences.

[0119]

[0120] S10.3: For B respectively norm (x i Perform first and second derivative calculations:

[0121]

[0122] Wherein, the sampling point interval, traversing the data points, if H(x i )·H(x i-1 If ) < 0, then it is determined that x is in the range of x. i With x i-1 If there is a zero-crossing point between them, they are denoted as potential edge point pairs;

[0123] S10.4: Calculate the mean μ of the magnitude of the first derivative. G and standard deviation σ G :

[0124]

[0125] Adaptive threshold T G =μ G +2σ G Only retain |G(x) i )|≥T G Zero-crossing pairs;

[0126] Each sub-interval contains only one defect. If there are multiple pairs of zero-crossing points, the pair with the largest gradient magnitude is selected as the candidate edge.

[0127] S10.5: For the filtered candidate edge pairs (x a ,x b ), in its 5 neighboring data points [x a-2 ,x a+2 Cubic spline interpolation is performed on the signal to obtain the high-density signal B. spline (x);

[0128] Fit a quadratic polynomial at three interpolation points near the zero-crossing point: B fit (x)=ax 2 +bx+c

[0129] Find the second derivative B″ fit (x) = 2a, and the zero point satisfies B″. fit (x) = 0, and the edge location is determined by combining the extreme points of the first derivative:

[0130] S10.6: Compare the results of the two determinations of the surface defect width, take the smaller value as the new edge of the surface defect, and then update the edge coordinates of the adjacent non-defect area.

[0131] This invention provides a comprehensive detection system and method for defects in steel plates. It offers the following advantages:

[0132] This comprehensive inspection system and method for steel plate defects organically integrates key subsystems such as image acquisition, magnetic flux leakage detection, and magnetic memory detection through modular and complete design. It is further enhanced by multi-level filtering and noise reduction and intelligent image processing algorithms, forming a highly industrialized and flexibly deployable inspection device. This device not only operates stably under complex working conditions but also enables rapid location and classification of internal and external defects in steel plates, significantly improving inspection accuracy and effectively reducing the risk of missed and false detections. This meets the stringent requirements for quality precision control in high-efficiency inspection work.

[0133] By fusing leakage magnetic field and magnetic memory detection data, and employing multiple filtering techniques to reduce noise in the magnetic field signal, the system performs multiple precise determinations of defect edges. Simultaneously, combined with visualized analysis of the overall stress state of the steel plate, a comprehensive assessment of the steel plate's health status under service conditions is achieved. This multimodal fusion and visualization assessment method not only improves the precision of defect detection but also provides an intuitive and reliable basis for subsequent maintenance decisions, further enhancing the intelligence level and application value of the entire detection system. Attached Figure Description

[0134] Figure 1 This is a flowchart illustrating a comprehensive testing method for realizing an invention;

[0135] Figure 2 This is a flowchart of the first method for implementing a comprehensive testing method for the invention;

[0136] Figure 3 This is a flowchart of the second method for implementing a comprehensive testing method of the invention;

[0137] Figure 4 This is a flowchart of the third method of a comprehensive testing method for realizing the invention;

[0138] Figure 5 This is an example diagram of a comprehensive testing system for implementing the invention;

[0139] Figure 6 This is an example diagram of another comprehensive testing system that realizes the invention.

[0140] The components include: 1. Front Hall sensor; 2. Rear Hall sensor; 3. Synchronous laser encoder; 4. Infrared sensor; 5. External power supply; 6. On-site large screen; 7. Computer terminal; 8. Mobile terminal; 9. Server; 10. Control device; 11. Steel plate to be tested; 12. Image acquisition device; 13. Semi-shielded active excitation device; 14. Liftable steering wheel assembly; 15. Lifting device. Detailed Implementation

[0141] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0142] like Figure 1-6 As shown, this embodiment of the invention provides a comprehensive detection system for steel plate defects, including an infrared sensor 4, a synchronous laser encoder 3, a Hall sensor, a control device 10, an image acquisition device 12, a semi-shielded excitation device, and a server 9.

[0143] Infrared sensor 4 is connected to control device 10. It detects whether the device has moved to the preset detection area of ​​the steel plate 11 to be detected and sends the detection result to control device 10. It is also used to detect whether the device has passed through the entire detection area. If the device has passed through the entire detection area, it sends the information that the device has passed through the entire detection area to control device 10.

[0144] The synchronous laser encoder 3 is connected to the control device 10. The encoder transmits the synchronous trigger time signal to the control device 10. When the encoder detection device moves to a position point, it sends a trigger signal to the controller. The image acquisition device 12 and the Hall sensor start acquiring data at the same time to ensure the consistency of data in time and position.

[0145] The Hall sensor is connected to the controller. When the Hall sensor receives the acquisition signal from the controller, it starts to acquire the magnetic field information of the current position and transmits the magnetic field information to the control device 10. The Hall sensor is placed on the lifting device 15.

[0146] The image acquisition device 12 is connected to the control device 10. When the image acquisition device 12 receives the acquisition signal from the control device 10, it starts the camera to acquire image information of the steel plate surface and sends it to the control device 10.

[0147] The control device 10 is connected to the infrared sensor 4, the synchronous laser encoder 3, the Hall sensor, the image acquisition device 12, and the server 9. It controls and manages each device, configures parameters, receives sensor detection results, and controls the start and stop of each part of the control device 10 based on the detection results. It is also responsible for transmitting the received data to the server 9. The neodymium iron boron permanent magnet weighs 1.8T. The integrated detection system also includes an outer frame support, which spans above the steel plate pretreatment line, providing support for the image acquisition device 12. The mechanical support is a frame structure with four adjustable steering wheel sets 14, and all devices are fixed to this outer frame. The laser emitter is fixed on the outer frame support and is used to project laser stripes of a specific frequency onto the upper surface of the steel plate. The illumination angle can be adjusted, and the height can also be adjusted by relying on the slide rails at both ends. The 3D industrial camera is deployed on the outer frame support, and its height can be adjusted by relying on the slide rails on both sides fixed to the outer frame to collect image information of the steel plate surface in the area illuminated by the laser emitter. The semi-shielded excitation device consists of a semi-shielded magnetic shielding barrel and an internal neodymium iron boron permanent magnet. This device is connected to the outer frame and can be raised and lowered freely. The neodymium iron boron permanent magnet is used to generate an excitation magnetic field that saturates the steel plate. The semi-shielded magnetic shielding barrel shields the outer edge interference, and most importantly, shields the excitation magnetic field generated by the device itself from interference with the rear Hall sensor 2.

[0148] The semi-shielded active excitation device 13 uses neodymium iron boron permanent magnets to saturate the detection area of ​​the steel plate, and the semi-shielded shell reduces the interference of the permanent magnets on the leakage magnetic field at the rear.

[0149] Server 9 receives the associated data uploaded by control device 10 and stores the associated data in three databases. It uses professional algorithms and models to conduct in-depth analysis of the data, identify the types and locations of defects on the steel plate surface, and perform location calibration. The results of defect detection are stored and fed back to relevant personnel so that timely measures can be taken to deal with them.

[0150] This embodiment also proposes a comprehensive detection method for steel plate defects, which specifically includes the following steps:

[0151] S1: Turn on the external power supply 5 to supply power to the entire set of equipment.

[0152] S2: The inspection trolley moves at a constant speed along the preset inspection line on the steel plate 11 to be inspected. When the infrared sensor 4 detects that it has reached the area to be inspected, it sends an arrival signal back to the control device 10.

[0153] S3: The control device 10 starts the synchronous laser encoder 3 based on the preset parameters and the system signal that it has entered the preset detection area, so that it outputs synchronous trigger signals at the set time intervals.

[0154] S4: When the synchronous trigger signal is issued, the control device 10 simultaneously triggers the front and rear Hall sensors 2 to collect magnetic field data, and the image acquisition device 12 to collect image data of the steel plate surface, and transmits the collected information to the controller in real time.

[0155] S5: The controller associates and packages time information, magnetic field data, and image data, and transmits the data collected at each time point to the server 9 wirelessly.

[0156] S6: Data acquisition ends when the rear infrared sensor 4 detects that the sensor has reached the edge of the detection area.

[0157] S7: Server 9 adjusts the actual location of the packaged information.

[0158] S8: Server 9 performs noise reduction processing on the magnetic field data collected by the front-end Hall sensor 1, then performs defect edge localization and non-defect region localization, and then proceeds to step S9. S8 includes the following steps:

[0159] Step S8.1: The unsaturated magnetized magnetic field strength data collected by the front Hall sensor 1 along the one-dimensional coordinate x is defined as a discrete sequence:

[0160] B = {B(x1), B(x2), ..., B(x...} N )}

[0161] Where x i =dx, where dx is the sampling point interval in mm, N is the total number of sampling points, and the detection range is x∈[0,L], L=N·dx.

[0162] For the defined discrete sequence, calculate the mean:

[0163]

[0164] Then the data is zero-mean normalized: B'(x) i )=B(x i )-μ, resulting in a new sequence B'={B'(x1),B'(x2),…,B'(x... N )}.

[0165] 8.2: Using the Mexican Hat Wavelet: Perform J-level discrete wavelet decomposition on B'(x) and use the Mallat algorithm to obtain the low-frequency approximate component A. J and high-frequency detail components D1, D2, ..., D J , corresponding to scale a j =2 j dx(j=1,2...J, e.g. J=5 corresponds to scales 1,2,4,8,16dx).

[0166] Then, the standard deviation of the noise is estimated using the first layer of high-frequency components:

[0167]

[0168] Where D1: the first layer of high-frequency detail components, median(|D1|): the median of the absolute values ​​of the elements in D1, and 0.6745: the conversion coefficient between the median and standard deviation of the Gaussian distribution.

[0169] For the high-frequency detail component D of each layer j Using unbiased likelihood estimation, the threshold λ for wavelet-based denoising is calculated:

[0170]

[0171] M = len(D) j ):D j The length of the high-frequency component in the j-th layer, d i :D j The i-th coefficient, sgn(d) i ): Sign function, 1 when di>0, 0 when di=0, -1 when di<0; Card counting function, counts the number of coefficients whose absolute value is greater than the threshold λ, iterates through λ≧0 and takes the λ that minimizes R(λ) as the threshold λ. SURE,j .

[0172] Soft thresholding is applied to the high-frequency component Dj in each layer. The formula for high-frequency component shrinkage is as follows:

[0173]

[0174] The wavelet is reconstructed layer by layer, and the formula is:

[0175]

[0176] S8.3: Calculate the wavelet coefficient modulus at various scales:

[0177]

[0178] a j =2 j ·dx: The scale corresponding to the j-th level decomposition, The Mexican hat wavelet after translation and scaling, and the signal in x i Convolution is performed at a specific location to extract mutation information at that scale.

[0179] For position x i If M(a) j ,x i )≥M(a j ,x i-1 And M(a)j ,x i )≥M(a j ,x i+1 If x i This represents the point where the modulus reaches its maximum at this scale, corresponding to the location of a sudden signal change.

[0180] From the largest scale a J Begin by recording the coordinates of all modulus maxima at this scale. For each At the sub-large scale a J-1 Search If a point with a modulus maximum exists within the range, retain that point. Repeat the above process until the smallest scale a1 is reached, ultimately retaining points that appear in ≥2 consecutive scales, i.e., points that simultaneously exist in scale a1. j and a j+1 scale.

[0181] S8.4: Sort the selected candidate points by their x-coordinates to form a sequence {X = x1, x2, ..., x_M}. Merge adjacent points with a distance less than or equal to 2dx to form clusters C1, C2, ..., C_M. k The mean of each cluster is used as the final edge coordinates:

[0182]

[0183] Specifically, regarding the determination of the left and right edges of the sub-defect: the abrupt change in wavelet coefficient of the magnetic field signal corresponding to the left edge from low to high changes from negative to positive. The abrupt change in wavelet coefficient of the magnetic field signal corresponding to the right edge from high to low changes from positive to negative, which serves as the calibration coordinate [x] for the edge of the defect region. left x right ].

[0184] S8.5: If K defects are detected, there are 2K edge coordinates x1. <x2<…<x 2k Then the non-defect area is:

[0185]

[0186] The non-defect region is divided into S according to the length ΔL. m =[x m ,x m +ΔL],x m This marks the starting point of the non-defect area.

[0187] For non-defect sub-intervals with intervals, magnetic characteristic parameters are selected, including:

[0188] Wavelet energy ratio:

[0189]

[0190] Among them, E(a) j ): Scale a j Wavelet energy in the non-defect region.

[0191] Signal variance:

[0192]

[0193] Where, N m :S m Inside points.

[0194] Modulus maximum density: K m For S m Number of internal modulus maxima, L m =ΔL, unit: mm.

[0195] For wavelet energy ratio R m Setting a threshold T under different stress states R1 ,T R2 ,T R3 Similarly, for the other two parameters, T is set... σ1 T σ2 ,T σ3 ,T D1 ,T D2 ,T D3 The threshold value was determined through standard specimen experiments.

[0196] Elastic phase: R m <T R1 , D m <T D1 .

[0197] Yield stage: T R1 ≤R m <T R2 , T D1 ≤D m <T D2 .

[0198] Enhancement Phase: T R2 ≤R m <T R3 , T D2 ≤ <T D3 .

[0199] Neck constriction phase: R m >T R3 , D m >T D3 .

[0200] If a certain subinterval S m If all three magnetic parameters satisfy a certain stress state threshold, the stress state is determined. If the stress state is different from the target stress state, a weighted method is used to assign weight values ​​to the three magnetic parameters, and the higher the weight, the higher the stress state. Based on the determination results of each sub-interval in the non-defect area, a different stress state is assigned to each interval.

[0201] S9: Server 9 selects the defect area for image processing, determines whether surface defects exist, and proceeds to step S10 after all sub-regions have been determined. S9 specifically includes the following steps:

[0202] 9.1: For the edge coordinates of the calibrated sub-defect region [x] left x right ], calculate the center position of the interval The center pixel coordinates of the image at the center position are (x center_px ,y center_px ), with x center_px Using the center as the reference point, select a rectangular region of size W*H as the Region of Interest (ROI), denoted as I. ROI .

[0203] S9.2 performs a single-channel grayscale conversion on the captured image. The formula for RGB to grayscale conversion is:

[0204] I(x,y)=0.299·R(x,y)+0.587·G(x,y)+0.114·B(x,y)

[0205] R(x,y), G(x,y), and B(x,y) are the range of three-channel pixel values ​​of the color image at position (x,y): 0 to 255, and I(x,y) is the converted grayscale value.

[0206] 9.3: Perform Gaussian filtering on the output grayscale image for noise reduction. Two-dimensional Gaussian distribution function:

[0207]

[0208] σ is the standard deviation, with a typical value of σ = 2.0 (the kernel size is an odd number, such as 5×5), i.e., x,y∈{-2,-1,0,1,2}.

[0209] For each pixel (i,j) in the image, calculate its weighted average with the Gaussian kernel:

[0210]

[0211] 9.4: For the ROI region after noise reduction, divide it into 8*8 sub-regions, count the gray values ​​of 0 to 255 of the 8×8=64 pixels one by one, record the number of times each gray level appears, and output an array h[i] of length 256, where h[i] is the number of pixels of the corresponding gray level and i is the gray level.

[0212] For the output array, set the clipLimit value to limit the histogram height, and the proportion of pixels of each gray level to the total number of pixels in the sub-region shall not exceed clipLimit / 255.

[0213] clipLimit: The probability density threshold, measured in pixels per gray level. It indicates the maximum number of pixels per gray level relative to the total number of pixels in a sub-region. If the number of pixels for a particular gray level in a sub-region exceeds clipLimit, the excess pixels are evenly distributed across all 256 gray levels. clipLimit is typically set to an empirical value of 1.0 to 3.0.

[0214] For all gray levels, the formula for calculating the cumulative distribution function is:

[0215]

[0216] clip(h(i),clipLimit): Limits the number of pixels for each gray level to no more than clipLimit, i.e., h(i) after clipLimit processing.

[0217] Perform grayscale mapping on each sub-region, assigning a new grayscale value to the pixels within that sub-region. The calculation formula is as follows:

[0218]

[0219] I blur (x,y): The grayscale value of this pixel after Gaussian filtering. CDF(I) blur (x,y): The cumulative number of pixels for this gray level.

[0220] Image I obtained after CLAHE processing clahe This is a single-channel grayscale image, where each pixel value represents the grayscale intensity at that location.

[0221] S9.5: Horizontal template G using the Sobel operator x and vertical template G y Used to detect edges in the horizontal and vertical directions respectively.

[0222]

[0223] For image I claheFor each pixel (i,j) in G, take its surrounding 3×3 neighborhood pixel block, and combine this neighborhood pixel block with G. x The elements at corresponding positions in the template are multiplied together, and then all products are summed to obtain the horizontal gradient value G of that pixel. x Repeat the above operation for each pixel (i,j) in the image to obtain the horizontal gradient image G. x Similarly, the vertical gradient image G is obtained. y .

[0224] For each pixel (i,j), calculate its gradient magnitude using the formula:

[0225]

[0226] Repeat the above operation for each pixel in the image to obtain the gradient magnitude image G.

[0227] For each pixel (i,j), calculate its gradient direction using the formula:

[0228]

[0229] The gradient direction is quantized to four main directions: 0°, 45°, 90°, and 135° for easier subsequent processing. The above operation is repeated for each pixel in the image to finally obtain the gradient direction image θ.

[0230] For each pixel (i,j) in the gradient magnitude image G, find its two neighboring pixels along the gradient direction θ(i,j), and compare the gradient magnitude G(i,j) of the pixel with the gradient magnitudes of these two neighboring pixels. If G(i,j) is less than the gradient magnitude of either neighboring pixel, set the gradient magnitude of the pixel to 0, indicating that the pixel is not an edge pixel. Otherwise, retain the gradient magnitude of the pixel. Repeat the above operation for each pixel in the image to obtain the gradient magnitude image G' after non-maximum suppression.

[0231] Based on the set high threshold T of the gradient magnitude h =150 and low threshold T l =50, for each pixel (i,j) in image G', if G'(i,j)≥T h If so, then the pixel is marked as a strong edge pixel. For example, T l ≤G'(i,j) <T h If T l If G'(i,j) is used, then the pixel is marked as a non-edge pixel.

[0232] Iterate through all pixels (i,j) marked as weak edges, and check the 8-neighborhood of each pixel.

[0233] If a strong edge pixel exists, mark the weak edge pixel as an edge pixel. Otherwise, mark it as a non-edge pixel.

[0234] For each pixel (i,j) in the image after double thresholding, if the pixel is marked as an edge pixel, set its pixel value to 255. Otherwise, set its pixel value to 0. Output the final binary edge map.

[0235] 9.6: For the output binary edge map, use OpenCV's findContours function to extract the continuous regions of all closed contour edges. Each contour is mapped to a coordinate list {(x1,y1),(x2,y2),…,(x...}. n ,y n )}.

[0236] For a closed contour to be defined, x must satisfy... n+1 =x1,y n+1 =y1, formula for contour area:

[0237]

[0238] Based on the actual minimum detectable size of the defect, a minimum area threshold A is set. threshold Values ​​less than this are considered noise.

[0239] Surface defect determination: If, among all extracted contours, there exists at least one contour with an area A ≥ A... threshold This is determined to be a surface defect. The area A of the contour... threshold It was determined to be an internal surface defect.

[0240] 9.7: Convert the grayscale ROI image to RGB format, use OpenCV's drawContours function to draw the contours, select red RGB value (255,0,0) as the color, set the line width to d=2 pixels, set the contour index to -1, draw all contours that meet the conditions, output a colored ROI image with red contour markers for surface defects, and add the label "Surface Defect", output the original grayscale ROI image for inner surface defects, and add the label "Inner Surface Defect".

[0241] S10: Server 9 expands the sub-regions of surface defects, and then extracts the magnetic field data corresponding to the expanded location region for new defect localization. S10 specifically includes the following steps:

[0242] ​S10.1 For image recognition, the defect coordinates [x] are determined to be surface defects. left ,x right Expanding the range at both ends by a percentage p% yields the target detection interval:

[0243]

[0244] Extract the rear Hall sensor 2 in [x left x right The magnetic field data sequence within the interval is B(x) = [B1, B2, ..., B...]. n ], where B i For position x i The magnetic field strength at that location.

[0245] 10.2: The magnetic field data is smoothed using a one-dimensional Gaussian kernel G(k,σ). The kernel function is:

[0246]

[0247] With a kernel size of 5 and a standard deviation σ = 1.0, the filtered data is as follows:

[0248]

[0249] The filtered data is mapped to the [0,1] interval to eliminate amplitude differences.

[0250]

[0251] S10.3: For B respectively norm (x i Perform first and second derivative calculations:

[0252]

[0253] Wherein, the sampling point interval, traversing the data points, if H(x i )·H(x i-1 If ) < 0, then it is determined that x is in the range of x. i With x i-1 There exists a zero point between them, denoted as a potential edge point pair x. i-1 x i .

[0254] S10.4: Calculate the mean μ of the magnitude of the first derivative. G and standard deviation σ G :

[0255]

[0256] Adaptive threshold T G =μ G +2σG Only retain |G(x) i )|≥T G The zero-crossing point pair.

[0257] Each sub-interval contains only one defect. If there are multiple pairs of zero-crossing points, the pair with the largest gradient magnitude is selected as the candidate edge.

[0258] S10.5: For the filtered candidate edge pairs (x a ,x b ), in its 5 neighboring data points [x a-2 ,x a+2 Cubic spline interpolation is performed on the signal to obtain the high-density signal B. spline (x).

[0259] Fit a quadratic polynomial at three interpolation points near the zero-crossing point: B fit (x)=ax 2 +bx+c

[0260] Find the second derivative B″ fit (x) = 2a, and the zero point satisfies B″. fit (x) = 0 is actually the point of sign change. The edge position can be determined by combining the extreme points of the first derivative: The extreme points of the first derivative, i.e., the edge locations.

[0261] S10.6: Compare the results of the two surface defect width determinations, take the smaller value as the new edge of the surface defect, and then update the edge coordinates of the adjacent non-defect area. Record the coordinates of the edge and width, as well as the corresponding magnetic field data, for different regions. The information storage for the defect area includes the image information corresponding to the defect center point. Map the edge coordinates of different intervals to the axial detection length, and mark the interval boundaries with line segments of different colors.

[0262] S11: Server 9 stores all information of the sub-intervals respectively, maps the edge coordinates of different intervals to the axial detection line, and marks the interval boundaries with line segments of different colors.

[0263] S12: Server 9 sends the steel plate defect location detection results to the on-site large screen 6 via wired communication. After receiving the detection information, the on-site large screen 6 displays it on-site, and on-site employees can observe the detected stress changes and defect location information of the steel plate through the on-site large screen 6.

[0264] S13: Employees log into the system using handheld mobile terminals 8, such as PAD devices. Based on the quantitative analysis information of steel plate defects stored in the host server 9, they can view the detailed situation of defect plate detection. According to the stress state distribution of the steel plate and the location of the steel plate defects, they can arrange personnel to carry out corresponding grinding, welding and other treatments.

[0265] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A comprehensive detection system for defects in steel plates, characterized in that, include: Infrared sensor (4), synchronous laser encoder (3), Hall sensor, control device (10), image acquisition device (12), semi-shielded excitation device and server (9); The infrared sensor (4) is connected to the control device (10); The synchronous laser encoder (3) is connected to the control device (10), and the encoder transmits the synchronous trigger time signal to the control device (10); The Hall sensor is connected to the controller. When the Hall sensor receives the acquisition signal from the controller, it starts to acquire the magnetic field information of the current position and transmits the magnetic field information to the control device (10). The image acquisition device (12) and the control device (10) are connected. When the image acquisition device (12) receives the acquisition signal from the control device (10), it starts the camera to acquire image information of the steel plate surface and sends it to the control device (10). The control device (10) is connected to the infrared sensor (4), the synchronous laser encoder (3), the Hall sensor, the image acquisition device (12), and the server (9), respectively. The semi-shielded active excitation device (13) uses neodymium iron boron permanent magnets to saturate the detection area of ​​the steel plate; The server (9) receives the associated data uploaded by the control device (10) and stores the associated data in three databases.

2. The comprehensive detection system for steel plate defects according to claim 1, characterized in that: The neodymium iron boron permanent magnet weighs 1.8T, and the integrated testing system also includes an outer frame support.

3. The comprehensive detection system for steel plate defects according to claim 2, characterized in that: The laser emitter is fixed on the outer frame support and is used to project laser stripes of a specific frequency onto the upper surface of the steel plate. The 3D industrial camera is deployed on the outer frame support.

4. The comprehensive detection system for steel plate defects according to claim 1, characterized in that: The semi-shielded excitation device consists of a semi-shielded magnetic shielding barrel and an internal neodymium iron boron permanent magnet.

5. A comprehensive detection method for defects in steel plates, characterized in that, Includes the following steps: S1: Turn on the external power supply (5) to supply power to the entire equipment; S2: The detection trolley moves at a constant speed along the preset detection line on the steel plate (11) to be detected. When the infrared sensor (4) detects that it has reached the area to be detected, it sends a signal of arrival to the control device (10). S3: The control device (10) starts the synchronous laser encoder (3) based on the preset parameters and the system has a signal that it has entered the preset detection area, so that it outputs a synchronous trigger signal according to the set time interval. S4: When the synchronous trigger signal is issued, the control device (10) simultaneously triggers the front and rear Hall sensors (2) to collect magnetic field data, and the image acquisition device (12) collects image data of the steel plate surface and transmits the collected information to the controller in real time. S5: The controller associates and packages the time information, magnetic field data, and image data, and transmits the data information collected at each time point to the server via wireless transmission (9); S6: When the rear infrared sensor (4) detects that it has reached the edge of the detection area, the data acquisition work ends; S7: The server (9) adjusts the actual location of the packaged information; S8: The server (9) performs noise reduction processing on the magnetic field data collected by the front Hall sensor (1), then performs defect edge positioning and non-defect interval positioning, and then proceeds to step S9. S9: The server (9) selects the image processing of the defect area to determine whether there is a surface defect. After all sub-intervals are determined, the process proceeds to step S10. S10: The server (9) expands the sub-intervals of the surface defects, and then extracts the magnetic field data corresponding to the expanded position intervals to locate the new defects. S11: The server (9) stores all the information of the sub-intervals respectively, maps the edge coordinates of different intervals to the axial detection line, and marks the interval boundaries with line segments of different colors; S12: The server (9) sends the steel plate defect location detection result information to the on-site large screen (6) through wired communication. After receiving the detection information, the on-site large screen (6) displays it on-site. On-site employees observe the detected stress change and defect location information of the steel plate through the on-site large screen (6). S13: Employees use handheld mobile terminals (8), such as PAD devices, to log in to the system. Based on the quantitative analysis information of steel plate defects stored in the host server (9), they can view the detailed situation of defect plate detection. According to the stress state distribution of the steel plate and the location of the steel plate defects, they can arrange personnel to carry out corresponding grinding, welding and other treatments.

6. The comprehensive detection method for defects in steel plates according to claim 5, characterized in that: S8 includes the following steps: Step S8.1: The unsaturated magnetized magnetic field strength data collected by the front Hall sensor (1) along the one-dimensional coordinate x is defined as a discrete sequence: B={B(x1),B(x2),…,B(x N )} Where x i =dx, where dx is the sampling point interval in mm, N is the total number of sampling points, and the detection range is x∈[0,L], L=N·dx; For the defined discrete sequence, calculate the mean: Then the data is zero-mean normalized: B ' (x i )=B(x i )-μ, to obtain a new sequence B ' ={B ' (x1),B ' (x2),…,B ' (x N )}; 8.2: Using the Mexican Hat Wavelet: For B ' (x) Perform J-level discrete wavelet decomposition and use the Mallat algorithm to obtain the low-frequency approximate component A. J and high-frequency detail components D1, D2, ..., D J , corresponding to scale a j =2 j dx(j=1,2...J, e.g. J=5 corresponds to scales 1,2,4,8,16dx); Then, the standard deviation of the noise is estimated using the first layer of high-frequency components: Where D1: the first layer of high-frequency detail components, median(|D1|): the median of the absolute values ​​of the elements in D1, and 0.6745: the conversion coefficient between the median and standard deviation of the Gaussian distribution; For the high-frequency detail component D of each layer j Using unbiased likelihood estimation, the threshold λ for wavelet-based denoising is calculated: M = len(D) j ):D j The length of the high-frequency component in the j-th layer, d i :D j The i-th coefficient, sgn(d) i ): Sign function, 1 when di>0, 0 when di=0, -1 when di<0; Card counting function, counts the number of coefficients whose absolute value is greater than the threshold λ, iterates through λ≧0 and takes the λ that minimizes R(λ) as the threshold λ. SURE,j ; Soft thresholding is applied to the high-frequency component Dj in each layer. The formula for high-frequency component shrinkage is as follows: The wavelet is reconstructed layer by layer, and the formula is: S8.3: Calculate the wavelet coefficient modulus at various scales: a j =2 j ·dx: The scale corresponding to the j-th level decomposition, The Mexican hat wavelet after translation and scaling, and the signal in x i Convolution is performed at a specific location to extract mutation information at that scale. For position x i If M(a) j ,x i )≥M(a j ,x i-1 And M(a) j ,x i )≥M(a j ,x i+1 If x i This represents the maximum modulus point at this scale, corresponding to the location of a signal abrupt change. From the largest scale a J Begin by recording the coordinates of all modulus maxima at this scale. For each At the sub-large scale a J-1 Search If a point with a modulus maximum exists within the range, retain that point; repeat the above process until the smallest scale a1 is reached, ultimately retaining points that appear in ≥2 consecutive scales, i.e., points that simultaneously exist in scale a1. j and a j+1 scale; S8.4: Sort the selected candidate points according to their x-coordinates to form a sequence {X = x1, x2, ..., x_M}; merge points whose adjacent distance is less than or equal to 2dx to form clusters C1, C2, ..., C_M. k The mean of each cluster is used as the final edge coordinates: Specifically, the determination of the left and right edges of the sub-defects involves: the left edge corresponding to a sudden change in the magnetic field signal from low to high; and the right edge corresponding to a sudden change in the magnetic field signal from high to low, which serves as the coordinate system for calibrating the edge of the defect region [x]. left x right ]; S8.5: If K defects are detected, there are 2K edge coordinates x1. <x2<…<x 2k Then the non-defect area is: The non-defect region is divided into S according to the length ΔL. m =[x m ,x m +ΔL],x m This marks the starting point of the non-defect area. For non-defect sub-intervals with intervals, magnetic characteristic parameters are selected, including: Wavelet energy ratio: Among them, E(a) j ): Scale a j Wavelet energy in the non-defect region; Signal variance: Where, N m :S m number of interior points; Modulus maximum density: K m For S m Number of internal modulus maxima, L m =ΔL, unit: mm; For wavelet energy ratio R m Setting a threshold T under different stress states R1 ,T R2 ,T R3 Similarly, for the other two parameters, T is set... σ1 T σ2 ,T σ3 ,T D1 ,T D2 ,T D3 The threshold value was determined through standard specimen experiments. Elastic phase: R m <T R1 , D m <T D1 ; Yield stage: T R1 ≤R m <T R2 , T D1 ≤D m <T D2 ; Enhancement Phase: T R2 ≤R m <T R3 , T D2 ≤ <T D3 ; Neck constriction phase: R m >T R3 , D m >T D3 .

7. The comprehensive detection method for defects in steel plates according to claim 5, characterized in that: S9 specifically includes the following steps: 9.1: For the edge coordinates of the calibrated sub-defect region [x] left x right ], calculate the center position of the interval The center pixel coordinates of the image at the center position are (x center_px ,y center_px ), with x center_px Using the center as the reference point, select a rectangular region of size W*H as the Region of Interest (ROI), denoted as I. ROI ; S9.2 performs a single-channel grayscale conversion on the captured image. The formula for RGB to grayscale conversion is: I(x,y)=0.299·R(x,y)+0.587·G(x,y)+0.114·B(x,y) R(x,y), G(x,y), B(x,y) are the three-channel pixel values ​​of the color image at position (x,y), and I(x,y) are the converted grayscale values. 9.3: Perform Gaussian filtering on the output grayscale image for noise reduction. Two-dimensional Gaussian distribution function: σ is the standard deviation, typical value: (σ=2.0) The kernel size takes an odd number (e.g. 5×5), that is, x,y∈{-2,-1,0,1,2}; For each pixel (i,j) in the image, calculate its weighted average with the Gaussian kernel: 9.4: For the ROI region after noise reduction, divide it into 8*8 sub-regions, count the gray values ​​(0~255) of 8×8=64 pixels one by one, record the number of times each gray level appears, and output an array h[i] of length 256, where h[i] is the number of pixels corresponding to the gray level and i is the gray level; For the output array, set the clipLimit value to limit the histogram height, and the proportion of pixels of each gray level to the total number of pixels in the sub-region shall not exceed clipLimit / 255; clipLimit: The threshold for probability density, measured in pixels per gray level. It indicates the maximum number of pixels per gray level. If the number of pixels for a gray level in a sub-region exceeds clipLimit, the excess pixels are evenly distributed across all 256 gray levels. clipLimit is typically set to 1.0–3.

0. For all gray levels, the formula for calculating the cumulative distribution function is: clip(h(i),clipLimit): Limits the number of pixels for each gray level to no more than clipLimit, i.e., h(i) after clipLimit processing; Perform grayscale mapping on each sub-region, assigning a new grayscale value to the pixels within that sub-region. The calculation formula is as follows: I blur (x,y): The grayscale value of this pixel after Gaussian filtering; CDF(I blur (x,y): The number of accumulated pixels for this gray level; Image I obtained after CLAHE processing clahe This is a single-channel grayscale image, where each pixel value represents the grayscale intensity at that location. S9.5: Horizontal template G using the Sobel operator x and vertical template G y Used to detect edges in the horizontal and vertical directions respectively. For image I clahe For each pixel (i,j) in G, take its surrounding 3×3 neighborhood pixel block, and combine this neighborhood pixel block with G. x The elements at corresponding positions in the template are multiplied together, and then all products are summed to obtain the horizontal gradient value G of that pixel. x Repeat the above operation for each pixel (i,j) in the image to obtain the horizontal gradient image G. x Similarly, the vertical gradient image G is obtained. y ; For each pixel (i,j), calculate its gradient magnitude using the formula: Repeat the above operation for each pixel in the image to obtain the gradient magnitude image G; For each pixel (i,j), calculate its gradient direction using the formula: The gradient direction is quantized to four main directions: 0°, 45°, 90°, and 135°. Based on the set high threshold T of the gradient magnitude h =150 and low threshold T l =50, for image G ' For each pixel (i,j) in the array, if G'(i,j)≥T h If T is an edge pixel, then the pixel is marked as a strong edge pixel; l ≤G'(i,j) <T h If T l If G'(i,j) is greater than the edge pixel, then the pixel is marked as a non-edge pixel. Iterate through all pixels (i,j) marked as weak edges, and check if there are any strong edge pixels in the 8-neighborhood of each pixel. If there are strong edge pixels, mark the weak edge pixel as an edge pixel; otherwise, mark it as a non-edge pixel. 9.6: For the output binary edge map, use OpenCV's findContours function to extract all closed contours, and map each contour to a coordinate list {(x1,y1),(x2,y2),…,(x... n ,y n )}; For a closed contour to be defined, x must satisfy... n+1 =x1,y n+1 =y1, formula for contour area: Based on the actual minimum detectable size of the defect, a minimum area threshold A is set. threshold Values ​​less than this are considered noise. 9.7: Convert the grayscale ROI image to RGB format, and use OpenCV's drawContours function to draw the outline, selecting red as the color (RGB value (255,0,0)).

8. A comprehensive detection method for defects in steel plates according to claim 5, characterized in that: S10 specifically includes the following steps: S10.1 For image recognition, the defect coordinates [x] are determined to be surface defects. left ,x right Expanding the range at both ends by a percentage p% yields the target detection interval: Extract the rear Hall sensor (2) in [x left x right The magnetic field data sequence within the interval is B(x) = [B1, B2, ..., B...]. n ], where B i For position x i Magnetic flux density at the location; 10.2: The magnetic field data is smoothed using a one-dimensional Gaussian kernel G(k,σ). The kernel function is: With a kernel size of 5 and a standard deviation σ = 1.0, the filtered data is as follows: The filtered data is mapped to the [0,1] interval to eliminate amplitude differences. S10.3: For B respectively norm (x i Perform first and second derivative calculations: Wherein, the sampling point interval, traversing the data points, if H(x i )·H(x i-1 If ) < 0, then it is determined that x is in the range of x. i With x i-1 There exists a zero point between them, denoted as a potential edge point pair (x). i-1 x i ); S10.4: Calculate the mean μ of the magnitude of the first derivative. G and standard deviation σ G : Adaptive threshold T G =μ G +2σ G Only retain |G(x) i )|≥T G Zero-crossing pairs; Each sub-interval contains only one defect. If there are multiple pairs of zero-crossing points, the pair with the largest gradient magnitude is selected as the candidate edge. S10.5: For the filtered candidate edge pairs (x a ,x b ), in its 5 neighboring data points [x a-2 ,x a+2 Cubic spline interpolation is performed on the signal to obtain the high-density signal B. spline (x); Fit a quadratic polynomial at three interpolation points near the zero-crossing point: B fit (x)=ax 2 +bx+c Find the second derivative B″ fit (x) = 2a, and the zero point satisfies B″. fit (x) = 0, and the edge location is determined by combining the extreme points of the first derivative: S10.6: Compare the results of the two determinations of the surface defect width, take the smaller value as the new edge of the surface defect, and then update the edge coordinates of the adjacent non-defect area.

Images