Cold-rolled sheet surface defect detection system based on visual analysis

Abstract

The application discloses a cold-rolled plate surface defect detection system based on visual analysis, relates to the technical field of surface defect detection, and mainly aims at the anti-interference detection problem of cold-rolled plates under complex warping working conditions. In the application, two subsystems run independently in parallel and complement each other, a thermal warping interference identification subsystem completes image interference elimination, region classification and pre-screening, a defect morphology identification subsystem completes deep identification, category determination and situation evaluation of defects, and a whole-process defect detection system of "interference pre-elimination, defect identification, situation dynamic evaluation and intelligent alarm by batches" is jointly constructed, the detection precision of cold-rolled plate surface defects under complex working conditions is effectively improved, the error alarm rate is reduced, data support is provided for rolling process optimization and defect traceability management, and the application is suitable for the industrialized high-speed online detection needs of various cold-rolled plates.

Description

Technical Field

[0001] This invention relates to the field of surface defect detection technology, specifically a surface defect detection system for cold-rolled steel sheets based on visual analysis. Background Technology

[0002] Cold-rolled sheet metal is a basic raw material for automobile manufacturing, precision hardware and other fields. Surface quality directly determines the appearance quality and mechanical performance of the end product. High-speed and high-precision surface defect detection is an indispensable quality control link in the cold-rolled production process.

[0003] Currently, traditional technologies lack regional differentiation in their processing methods, applying uniform detection parameters to warped, severely distorted, and normal board surface areas. This leads to a series of problems, such as undetected defects in ambiguous areas, false detections in virtual image areas, and redundant detection in normal areas. It cannot achieve anti-interference detection under complex warping conditions. Furthermore, it lacks a quantitative evaluation system for defect evolution, making it difficult to determine the growth, deformation, and attenuation states of defects, thus failing to distinguish between continuously deteriorating high-risk primary defects and process-controllable secondary defects, resulting in undifferentiated defect alarms.

[0004] To address the aforementioned technical shortcomings, a visual analysis-based surface defect detection system for cold-rolled steel sheets is proposed. The overall architecture consists of two parallel operating units: a thermal warpage interference recognition subsystem and a defect morphology identification subsystem. Designed for high-speed online real-time detection scenarios of cold-rolled steel sheets, this system addresses industry pain points such as extreme thermal warpage interference, indistinguishable defect categories, and rigid detection parameters by constructing an intelligent defect detection system with temporal evolution analysis and adaptive threshold adjustment. Summary of the Invention

[0005] The purpose of this invention is to solve the problems mentioned above by proposing a visual analysis-based surface defect detection system for cold-rolled sheets.

[0006] The objective of this invention can be achieved through the following technical solution: a surface defect detection system for cold-rolled sheet metal based on visual analysis, including a thermal warpage interference identification subsystem and a defect morphology discrimination subsystem; The thermal warpage interference identification subsystem collects data sources, which are divided into bright field illumination and structured light illumination. Based on different data sources, it collects features, extracts the surface area of ​​the cold-rolled sheet and divides it into several sub-regions, and determines the features of each sub-region, specifically: the local curvature change rate of the laser line, the local energy attenuation rate of the laser line, and the angle deviation between the local gray-scale gradient direction and the warpage normal. Alarm decisions are made based on feature comparison. The defect morphology identification subsystem denoises and aligns N consecutive frames of surface images, locates the defect region in each frame, outputs a multi-frame defect local image dataset, extracts defect edge coordinates and texture entropy values ​​frame by frame, collects the temporal diffusion gradient of defect morphology and the temporal entropy decay rate of texture, calculates the diffusion gradient and entropy decay rate to generate a defect temporal evolution trend index sequence, retrieves a pre-constructed cold rolling defect temporal evolution standard database, dynamically adjusts the defect judgment threshold based on the deviation between the frame mean of the real-time evolution trend index sequence and the standard benchmark value, sets defects exceeding the adjusted threshold as potential defects, extracts static morphological feature vectors and temporal evolution feature vectors, calculates the orthogonal matching degree of morphology-evolution features, and performs corresponding defect alarm maintenance according to the matching degree interval.

[0007] Furthermore, the bright field lighting data source is: acquiring the original grayscale image I(x,y) of the board surface, where x is along the board width direction (lateral) and y is along the rolling direction (longitudinal), and setting it as a bright field image; Structured light illumination: A series of continuous laser lines parallel to the rolling direction are projected at a shallow angle, and a structured light image L(x,y) is acquired.

[0008] Furthermore, the local curvature change rate of the laser line is extracted. For each laser line, its centerline coordinate c(y) is extracted, and the first derivative C′(y) and second derivative c"(y) are calculated. The sign oscillation exponent of the curvature change rate is then further calculated. ; The subscript k is used to distinguish different points; The third derivative of function c with respect to y at y=y k The value at; The fourth derivative of function c with respect to y in y=y k The value at; "and": requires both conditions to be met simultaneously; =0 and ≠0: respectively, meaning "equal to zero" and "not equal to zero".

[0009] Furthermore, the local energy attenuation rate of the laser line is extracted, and the gray-scale integral of the light stripe cross-section is obtained in a small neighborhood around the laser line and compared with the integral value obtained during calibration. ; Among them, L calib (x,y) is the reference signal used for calibration; c0(y) is the calibrated ideal center line, and δ is dynamically determined based on the half-height and full width of the measured laser line width.

[0010] Furthermore, the angle deviation between the local gray-level gradient direction and the warping normal is extracted from the bright-field illumination data source, and the gray-level gradient direction θ of the candidate edge points is calculated. I(x,y) and the normal gradient direction θ of the warped surface reconstructed by the laser line. S(x,y) And define the included angle difference: .

[0011] Furthermore, the feature comparison alarm decision-making steps are as follows: If the rate of change of local curvature of the laser line in a sub-region exceeds the rate of change threshold, the current sub-region is set as a warped region; if the rate of change of local curvature of the laser line in a sub-region does not exceed the rate of change threshold, it is set as a non-warped region. If the local energy attenuation rate of the laser line in a sub-region is less than the attenuation rate threshold, it is set as a high blur region. If the angle deviation between the local gray-level gradient direction of the sub-region and the warping normal does not exceed the angle deviation threshold, the corresponding sub-region is set as a defect candidate region, retained, and output. Based on low-fuzzy regions and defect candidate regions, warped regions and non-warped regions are screened out to determine the distribution location of warped regions in cold-rolled sheets. Alarm decisions are then made based on the quality requirements of cold-rolled sheets in conjunction with the processing technology.

[0012] Furthermore, the formula for calculating the temporal diffusion gradient of the defect morphology is as follows: ; Where M: total number of time-series images, S k : Area enclosed by the defect contour in the k-th frame; Δt: Time interval between adjacent frames; The formula for calculating the temporal entropy decay rate of texture is: Where E1: texture information entropy of the defect region in the first frame, and EM: texture information entropy of the defect region in the last frame.

[0013] Furthermore, when the defect temporal evolution trend index exceeds the defect judgment threshold, it is set as a potential defect, and the orthogonal matching degree of morphological-evolutionary features is calculated.

[0014] Furthermore, the formula for calculating the orthogonal matching degree of morphological-evolutionary features is: V sta : Static morphological feature vector of defects (area, contour, grayscale distribution, etc.) V evo : Defect temporal evolution feature vector, ‖·‖: Vector L2 norm.

[0015] Furthermore, if the orthogonal matching degree of morphological-evolutionary features falls within the standard range of primary defects in the database, it is set as a high-risk primary defect; if the orthogonal matching degree of morphological-evolutionary features falls within the range of secondary defects in the database, it is set as a process-induced derivative defect.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Compared to traditional technologies, the thermal warping interference identification subsystem possesses multi-dimensional technical advantages and engineering application value. Firstly, it innovatively constructs a bright-field + shallow-angle structured light dual-imaging acquisition system, overcoming the information limitations of a single bright-field image. By reconstructing the three-dimensional warping morphology of the board material through structured light images, it uncovers three types of interference features caused by warping: distortion, defocus attenuation, and specular reflection virtual images. For the first time, it achieves decoupling analysis of the nonlinear coupling relationship between warping morphology and image distortion, solving the problem that traditional technologies cannot distinguish between virtual images and defects. Secondly, it abandons the traditional high-damage, high-cost hardware flattening scheme and the fixed-parameter coarse-grained software compensation scheme. Through multi-feature threshold comparison, it achieves refined classification of board material regions, dividing them into warped / non-warped areas, high-blurred / low-blurred areas, and virtual image interference / defect candidate areas. Different fuzzy correction parameters and defect detection standards are applied to different areas, significantly reducing the undetected rate and false detection rate under extreme conditions.

[0017] The defect morphology identification subsystem and the thermal warping interference recognition subsystem operate in parallel and collaboratively to achieve in-depth defect identification and targeted control, overcoming the shortcomings of traditional detection technologies such as low classification capabilities, single evaluation dimensions, and rigid parameters. First, it innovatively introduces implicit temporal evolution analysis, breaking through the limitations of single-frame static detection. By using the temporal diffusion gradient of defect morphology and the temporal entropy decay rate of texture, it constructs a defect temporal evolution trend index, quantifying the dynamic evolution state of defects and filling the technological gap in dynamic evaluation of cold-rolled defects in the industry. Second, it constructs a working condition adaptive threshold adjustment mechanism. Relying on a massive working condition standard database, it dynamically adjusts the defect judgment threshold by mapping the deviation between real-time defect evolution characteristics and standard benchmark values, completely solving the problem that traditional fixed thresholds cannot adapt to working condition fluctuations and significantly improving the defect detection accuracy under complex working conditions. Attached Figure Description

[0018] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0019] Figure 1 This is a system principle block diagram of the present invention; Figure 2 This is a flowchart of the thermal warp interference identification subsystem in this invention; Figure 3 This is a flowchart of the defect morphology identification subsystem in this invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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.

[0021] In this document, the term "embodiment" means that the structure or feature described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0022] Please see Figures 1-3 As shown, the surface defect detection system for cold-rolled sheet metal based on visual analysis is divided into a thermal warping interference recognition subsystem and a defect morphology identification subsystem. The two subsystems are used for recognition processing to complete the defect detection system. The thermal warpage interference recognition subsystem is applicable to scenarios where thermal warpage occurs in cold-rolled steel sheets. Existing visual inspection systems assume the sheet material is an ideal plane and use a line scan camera with a bright field light source to acquire images. For edge areas, simple "grayscale threshold segmentation" or "edge contour tracking" is typically used to identify defects (such as edge cracks and burrs). When warpage exists, the system either forces flattening through mechanical rollers or air blowing devices (increasing hardware costs and easily damaging the sheet surface), or treats the defocusing blur and distortion caused by warpage as abnormal areas and compensates for them through fixed image enhancement (such as histogram equalization). The fundamental flaw in the above approach is that it treats "board material (warping)" and "surface defects" as two independent problems, employing a linear pipeline of "first acquiring images → then simple correction." However, in practice, warping not only causes non-uniform distortion and spatial defocusing of the image, but more seriously, the warped area changes the specular reflection direction of the light source, producing "virtual images"—the background that should be flat appears with pits or bumps, creating error textures. Conventional grayscale correction or homography transformation cannot distinguish between virtual images and defects because virtual images cannot be separated from true defects in grayscale statistics. Therefore, the thermal warpage interference recognition subsystem solves the problem that the nonlinear coupling relationship between warpage morphology and image distortion is ignored, which leads to the inability to distinguish virtual images / defocus blur and surface defects caused by thermal warpage under conditions of large temperature difference at the edges. During the cold-rolled sheet processing stage, the surface of the sheet is inspected in real time; The data source acquisition device consists of a single linear scan camera, coupled with two imaging channels (using either beam splitting or sequential alternating illumination): Data sources are divided into: Bright field lighting: Acquire the original grayscale image I(x,y) of the plate surface, where x is along the plate width direction (lateral) and y is along the rolling direction (longitudinal), and set it as a bright field image; Structured light illumination: Project a series of continuous laser lines parallel to the direction (line spacing 10~20mm) at a shallow angle (15°~30°), and acquire structured light images L(x,y).

[0023] Feature acquisition of structured light images: Extract the local curvature change rate of the laser line. For each laser line, extract its centerline coordinate c(y), calculate the first derivative C′(y) and the second derivative c"(y), and further calculate the sign oscillation exponent of the curvature change rate: ; Within a sliding window (the window length corresponds to the expected length of a warped half-wave), count the number of zero crossings of c"(y). Meaning: a high number of zero crossings indicates a steep warped waveform (rich in higher harmonics), and the corresponding image distortion is more severe.

[0024] Among them, y k A variable with a subscript k represents all possible points of the independent variable (usually points within the domain of the function c(y)). The subscript k is used to distinguish different points.

[0025] The third derivative of function c with respect to y at y=y k The value at; The fourth derivative of function c with respect to y in y=y k The value at; "and": requires both conditions to be met simultaneously; =0 and ≠0: represent "equal to zero" and "not equal to zero" respectively; Simultaneously extract the local energy attenuation rate of the laser line, calculate the gray-scale integral of the light stripe cross section within a small neighborhood around the laser line, and compare it with the integral value during calibration (on a flat plate surface): ; Among them, L calib (x,y) is the reference signal used for calibration. It is usually a signal measured under standard conditions (such as uniform illumination and a clear target) and is used to eliminate the effects of detector non-uniformity. c0(y) is the calibrated ideal center line, and δ is dynamically determined based on the half-height and full width of the measured laser line width. Feature extraction from bright-field images; Extract the angular deviation between the local gray-level gradient direction and the warped normal, and calculate the gray-level gradient direction θ of the candidate edge points. I(x,y) and the normal gradient direction θ of the warped surface reconstructed by the laser line. S(x,y) Then define the angle difference: ; Meaning: The Δθ of the defect is uniformly distributed between 0 and 360°, while the Δθ of the virtual image has a bimodal distribution (concentrated at 0° and 180°). The surface area of ​​the cold-rolled sheet is extracted and divided into several sub-regions. The local curvature change rate, local energy attenuation rate, and angle deviation between the local gray-scale gradient direction and the warping normal are determined for each sub-region. If the local curvature change rate of the laser line in a sub-region exceeds the change rate threshold, it is determined that there is a "change" at that location, which may correspond to defects such as cracks, pits, or bumps. It needs to be analyzed, and the current sub-region is set as a warped area. If the local curvature change rate of the laser line in a sub-region does not exceed the change rate threshold and the warped waveform is smooth, it belongs to a large-scale material problem rather than a local defect. It can be processed in the usual way during correction and set as a non-warped area. If the local energy attenuation rate of the laser line in the sub-region is less than the attenuation rate threshold, the region is determined to be "severely out of focus". A larger blur kernel is used when deblurring, and the contrast value requirement is reduced when performing defect detection in this region (because blur may remain even after correction), and it is set as a high blur region. If the local energy attenuation rate of the laser line in the sub-region is not less than the attenuation rate threshold, it will be processed according to normal parameters and set as a low-blur region. If the angle deviation between the local gray-level gradient direction and the warping normal of a sub-region exceeds the angle deviation threshold, the corresponding sub-region will be set as a virtual image and directly removed without alarm. If the angle deviation between the local gray-level gradient direction of the sub-region and the warping normal does not exceed the angle deviation threshold, the corresponding sub-region is set as a defect candidate region, retained, and output. Based on low-fuzzy regions and defect candidate regions, warped regions and non-warped regions are screened out. That is, when low-fuzzy regions overlap with warped regions, region analysis is performed again, or when defect candidate regions overlap with warped regions, region categories are determined. The location of the warped area of ​​the cold-rolled sheet is determined, and an alarm decision is made based on the quality requirements of the cold-rolled sheet in combination with the processing technology. If the area of ​​the warped area does not affect the quality of the actual acquired image or the area of ​​the warped area is within the processing error range, no alarm is issued. Otherwise, an alarm is issued and the cold-rolled sheet is reworked. The defect morphology identification subsystem performs image recognition and monitoring on cold-rolled sheets based on images collected through visual analysis. The current subsystem and the thermal warping interference identification subsystem are executed in parallel, meaning that the defect area of ​​the current subsystem does not cover the warping area. Using a sequence of N consecutive cold-rolled sheet surface images acquired in real time from the production line as input, the image sequence is denoised and aligned to eliminate inter-frame offset errors caused by sheet vibration and light and shadow interference. Defect regions in each frame are located through static contour features, and a multi-frame defect local image dataset with time-series labels is output. Based on the output temporal defect image dataset, the defect edge coordinates and texture entropy values ​​are extracted frame by frame, and the latent features of defect morphology temporal diffusion gradient and texture temporal entropy decay rate are calculated respectively. Among them, the temporal diffusion gradient of the defect morphology is used to reflect the deformation expansion rate of the defect edge with the rolling time sequence. It is obtained by statistically analyzing the change in the defect contour area per unit time within consecutive frames, specifically: ; Where M: total number of time-series images, S k : Area enclosed by the defect contour in the k-th frame; Δt: Time interval between adjacent frames. The texture temporal entropy decay rate is used to reflect the degree of fading and decay of texture details in defective areas as the process progresses. It is obtained by statistically analyzing the decay ratio of defect texture entropy values ​​between the first and last frames, specifically: Where E1: texture information entropy of the defect region in the first frame, and EM: texture information entropy of the defect region in the last frame; After obtaining the latent temporal features, the features are normalized and weighted to generate an intermediate abstract index, the defect temporal evolution trend index. α, β: normalized weights, α+β=1, adaptively set according to working conditions; It is used to quantify the comprehensive state of defect growth, deformation, and decay, and finally outputs the evolution trend index sequence of each defect under continuous time series. The weight coefficients satisfy the normalization constraint conditions and can be adaptively and dynamically adjusted according to the real-time rolling conditions.

[0026] The system retrieves a locally pre-built standard database of cold-rolled defect time-series evolution. This database stores standard intervals of evolutionary trend indices for primary and secondary defects under different rolling speeds and plate thicknesses. Using the obtained real-time evolutionary trend index sequence as input, the system dynamically adjusts the defect judgment threshold for the current detection scenario based on the deviation mapping relationship between the real-time feature mean and the standard benchmark value, thus upgrading from fixed parameters to adaptive dynamic parameters. Specifically: ; T adj : Defect judgment threshold after adaptive adjustment; Tbase : Default threshold for basic operating conditions; λ: Dynamic modulation coefficient; : The frame mean of the current defect temporal evolution trend index; I std : The benchmark value of the standard evolution index for database storage under the corresponding operating conditions; When the defect evolution trend index exceeds the defect judgment threshold, it is set as a potential defect and positive interaction verification is performed. If the defect time-series evolution trend index does not exceed the defect judgment threshold, it is regarded as background noise and directly filtered out; Extract the static morphological features (area, contour, grayscale distribution) of the defect region in the current frame to construct a static morphological feature vector. Combine the output adaptive dynamic judgment parameters to construct a temporal evolution feature vector, forming two sets of mutually independent constraint spaces to achieve positive interactive verification and obtain the orthogonal matching degree of morphological-evolutionary features. ; V sta : Static morphological feature vector of defects (area, contour, grayscale distribution, etc.) V evo : Defect temporal evolution feature vector; ‖·‖: Vector 2 norm; If the orthogonal matching degree of morphological-evolutionary features falls within the original defect standard range of the database, it is set as a high-risk original defect, and targeted defect alarm maintenance is performed based on the high-risk original defect. If the orthogonal matching degree of morphological-evolutionary features falls within the secondary defect range of the database, it is set as a process-induced derivative defect, and defect alarm maintenance is performed based on the process-induced derivative defect. High-risk defects originating from the material or the initial stage of the rolling process refer to inherent defects that already exist, such as surface microcracks, inclusions inside the incoming slab, and original pores in the base material. These defects often deteriorate spontaneously over time and are highly detrimental to product performance, requiring monitoring. Secondary derivative defects refer to surface damage induced by changes in process conditions (such as tension fluctuations, uneven cooling, and vibration) during rolling. Examples include scratches (in contact with guides and rollers), vibration marks, and thermal scratches (localized overheating friction). These defects are usually strongly correlated with rolling parameters, do not necessarily continue to grow, and some can be suppressed by adjusting the process.

[0027] By comparing with the current subsystems, defect maintenance can be more targeted and completed more efficiently.

[0028] In summary, the thermal warpage interference identification subsystem can identify minute defects in high-strength steel thin strips under large temperature difference rolling conditions, thus solving the problem of undetected microcracks caused by defocusing blur in traditional algorithms. At the same time, it batch-eliminates massive error alarms caused by periodic specular reflection virtual images, greatly improving the accuracy of the detection system. In addition, the differentiated processing effectively adapts to different degrees of sheet distortion problems. Lightweight correction processing is used for common sheet problems with large-scale smooth warpage, and defect risk areas with local changes are detected. While ensuring detection accuracy, it effectively reduces system computational redundancy, improves online real-time detection efficiency, and is fully compatible with the high-speed continuous production rhythm of cold rolling production lines.

[0029] The defect morphology identification subsystem innovatively adopts an orthogonal matching verification mechanism combining static morphological features and temporal evolution features. Through independent cross-validation of these two dimensions, it reduces the probability of misjudgment based on single-feature detection, significantly improving defect identification accuracy. It distinguishes between primary high-risk defects and secondary process defects. For primary defects with significant spontaneous hazards, it triggers alarms and rectification maintenance; for process-induced and controllable secondary defects, it provides process optimization directions, achieving differentiated defect alarms and avoiding ineffective maintenance. Furthermore, this subsystem enables continuous tracking of defect evolution, providing comprehensive data support for rolling equipment fault diagnosis, process parameter optimization, and product quality traceability. This helps cold rolling production lines upgrade from "passive defect detection" to "proactive quality control," effectively improving the finished product qualification rate and the level of refined production management.

[0030] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and changes can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A visual analysis-based surface defect detection system for cold-rolled sheet metal, characterized in that, This includes a thermal warpage interference identification subsystem and a defect morphology identification subsystem; The thermal warpage interference identification subsystem collects data sources, which are divided into bright field illumination and structured light illumination. Based on different types of data sources, it collects features, extracts the surface area of ​​the cold-rolled sheet and divides it into several sub-regions, and determines the features of each sub-region, specifically: the local curvature change rate of the laser line, the local energy attenuation rate of the laser line, and the angle deviation between the local gray-scale gradient direction and the warpage normal. Alarm decisions are made based on feature comparison. The defect morphology identification subsystem locates the defect region, obtains a multi-frame local image dataset of the defect, extracts the defect edge coordinates and texture entropy values, and calculates the temporal diffusion gradient of the defect morphology and the temporal entropy decay rate of the texture. The formula for calculating the temporal diffusion gradient of the defect morphology is as follows: ； Where M: total number of time-series images, S k : Area enclosed by the defect contour in the k-th frame; Δt: Time interval between adjacent frames; The formula for calculating the temporal entropy decay rate of texture is: Where E1: texture information entropy of the defect area in the first frame, EM: texture information entropy of the defect area in the last frame, to obtain the defect temporal evolution trend index sequence, adjust the defect judgment threshold, obtain potential defects, extract static morphological feature vector and temporal evolution feature vector, calculate the orthogonal matching degree of morphological-evolutionary features, and perform corresponding defect alarm maintenance according to the matching degree interval.

2. The visual analysis-based surface defect detection system for cold-rolled sheet metal according to claim 1, characterized in that, Bright field lighting data source: Obtain the original grayscale image I(x,y) of the board surface, where x is along the board width direction (lateral) and y is along the rolling direction (longitudinal), and set it as a bright field image; Structured light illumination: A series of continuous laser lines parallel to the rolling direction are projected at a shallow angle, and a structured light image L(x,y) is acquired.

3. The visual analysis-based surface defect detection system for cold-rolled sheet metal according to claim 2, characterized in that, Extract the local curvature change rate of the laser line. For each laser line, extract its centerline coordinate c(y), calculate the first derivative C′(y) and the second derivative c"(y), and further calculate the sign oscillation exponent of the curvature change rate: ； The subscript k is used to distinguish different points; The third derivative of function c with respect to y at y=y k The value at; The fourth derivative of function c with respect to y in y=y k The value at; "and": requires both conditions to be met simultaneously; =0 and ≠0: respectively, meaning "equal to zero" and "not equal to zero".

4. The visual analysis-based surface defect detection system for cold-rolled sheet metal according to claim 3, characterized in that, Extract the local energy attenuation rate of the laser line, calculate the gray-scale integral of the light stripe cross-section in a small neighborhood around the laser line, and compare it with the integral value during calibration: ； Among them, L calib (x,y) is the reference signal used for calibration; c0(y) is the calibrated ideal center line, and δ is dynamically determined based on the half-height and full width of the measured laser line width.

5. The visual analysis-based surface defect detection system for cold-rolled sheet metal according to claim 4, characterized in that, The angle deviation between the local gray-level gradient direction and the warping normal is extracted from the bright-field lighting data source, and the gray-level gradient direction θ of the candidate edge points is calculated. I(x,y) and the normal gradient direction θ of the warped surface reconstructed by the laser line. S(x,y) And define the included angle difference: 。 6. The visual analysis-based surface defect detection system for cold-rolled sheet metal according to claim 5, characterized in that, The feature comparison alarm decision-making steps are as follows: If the rate of change of local curvature of the laser line in a sub-region exceeds the rate of change threshold, the current sub-region is set as a warped region; if the rate of change of local curvature of the laser line in a sub-region does not exceed the rate of change threshold, it is set as a non-warped region. If the local energy attenuation rate of the laser line in a sub-region is less than the attenuation rate threshold, it is set as a high blur region. If the angle deviation between the local gray-level gradient direction of the sub-region and the warping normal does not exceed the angle deviation threshold, the corresponding sub-region is set as a defect candidate region, retained, and output. Based on low-fuzzy regions and defect candidate regions, warped regions and non-warped regions are screened out to determine the distribution location of warped regions in cold-rolled sheets. Alarm decisions are then made based on the quality requirements of cold-rolled sheets in conjunction with the processing technology.

7. The visual analysis-based surface defect detection system for cold-rolled sheet metal according to claim 1, characterized in that, When the defect temporal evolution trend index exceeds the defect judgment threshold, it is set as a potential defect, and the morphological-evolutionary feature orthogonal matching degree is used.

8. The visual analysis-based surface defect detection system for cold-rolled sheet metal according to claim 7, characterized in that, The formula for the orthogonal matching degree of morphological-evolutionary features is: V sta : Defect static morphological feature vector; V evo : Defect temporal evolution feature vector, ‖·‖: Vector L2 norm.

9. The visual analysis-based surface defect detection system for cold-rolled sheet metal according to claim 8, characterized in that, If the orthogonal matching degree of morphological-evolutionary features falls within the standard range of primary defects in the database, it is set as a high-risk primary defect. If the orthogonal matching degree of morphological-evolutionary features falls within the range of secondary defects in the database, it is set as a process-induced derivative defect.

Images