Image recognition-based online detection method and system for surface defects of ultra-wide and thick plates

Abstract

The application provides an image recognition-based online detection method and system for surface defects of ultra-wide and thick plates, and relates to the technical field of image recognition. The method comprises the following steps: synchronously collecting surface images and process state parameters containing final rolling temperature, rolling force distribution, laminar cooling intensity and real-time speed, firstly completing image splicing according to the final rolling temperature and the rolling force distribution to obtain a topological image; then separating a texture channel image by feature decoupling with the laminar cooling intensity as a constraint; subsequently extracting edge features of the texture channel image by Hough transformation, and introducing the rolling force distribution and the real-time speed as constraints for curve fitting to generate topological structure features; finally dynamically adjusting a kernel function by a support vector machine, and analyzing the topological structure features and a preset feature library to obtain a defect detection result. The application can effectively improve the accuracy and adaptability of defect detection under complex working conditions through deep fusion and dynamic guidance of process parameters.

Description

Technical Field

[0001] This application relates to the technical field of image recognition, and in particular to an online detection method and system for surface defects of ultra-wide and thick plates based on image recognition. Background Technology

[0002] In key industrial sectors such as steel, shipbuilding, and heavy equipment manufacturing, ultra-wide and thick plates are core structural materials. Their surface quality directly affects the safety and performance of the final product. Furthermore, using visual methods to automatically inspect the surface of plates on the production line is of great application value for timely detection of defects and ensuring product quality.

[0003] Currently, the testing equipment deployed on the production site usually executes a fixed testing procedure. After taking pictures of the board, these devices analyze the pictures according to preset judgment rules, such as fixed thresholds based on image grayscale or contrast, to determine whether there are abnormal areas.

[0004] However, in actual production, the surface condition of the board will vary significantly with the real-time changes in process parameters. The fixed program of the existing detection device cannot sense and adapt to the changes in imaging conditions caused by process fluctuations. It is easy to misjudge normal texture fluctuations caused by the process as defects or miss real defects, which leads to an increase in the false alarm rate and false alarm rate of the detection results when production conditions change. Summary of the Invention

[0005] The purpose of this application is to provide an online detection method and system for surface defects of ultra-wide and thick plates based on image recognition, so as to solve the problem that the reliability of detection results is greatly affected by real-time production process fluctuations in the existing technology.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides an online detection method for surface defects in ultra-wide and thick plates based on image recognition, comprising: Surface images and process parameters of ultra-wide and thick plates are collected from multiple locations. The process parameters include final rolling temperature, rolling force distribution, laminar cooling intensity, and real-time speed. Based on the final rolling temperature and rolling force distribution, all the surface images are stitched together to obtain a topological image of the ultra-wide and thick plate synchronized with the process. The topological image is input to a physically guided feature decoupling module. The feature decoupling module, constrained by the laminar cooling intensity, decomposes the topological image into a texture channel image representing defects and a background channel image representing the substrate material and oxide layer using a reflectivity separation algorithm. The texture channel image is processed by Hough transform to obtain edge features. The rolling force distribution and real-time speed are used as regularization constraints to perform curve fitting on the edge features to generate topological features. The topological features and a preset feature library are analyzed using a support vector machine. The kernel function of the support vector machine is selected by dynamically adjusting the final rolling temperature and laminar cooling intensity to obtain defect detection results.

[0007] Optionally, the analysis of the topological features and a preset feature library using a support vector machine (SVM) is performed, wherein the kernel function of the SVM is dynamically adjusted based on the final rolling temperature and laminar cooling intensity to obtain defect detection results, including: Based on the final rolling temperature and laminar cooling intensity, a preset process condition mapping table is consulted to determine the condition code to which the current state belongs. Using a support vector machine, the linear kernel and radial basis kernel of the support vector machine are dynamically combined according to the kernel function configuration scheme corresponding to the working condition code, and the parameters of each kernel are set to construct a hybrid kernel function. Based on the hybrid kernel function, the similarity between the topological structure features and the feature templates in the preset feature library is calculated to obtain the similarity distribution. The similarity distribution is subjected to high-dimensional space mapping to obtain classification decision features; A classification mechanism is used to perform pattern recognition processing on the classification decision features to obtain the defect type. Based on the defect type, combined with the defect geometric information parsed from the topological structure features, the size and location coordinates of the defect are determined. The defect type, size, and location coordinates are fused together to generate a defect detection result.

[0008] Optionally, the step of calculating the similarity distribution between the topological features and feature templates in a preset feature library based on the hybrid kernel function includes: Based on the hybrid kernel function, the high-dimensional similarity between the topological structure features and the feature templates of various defects in the preset feature library is calculated to obtain an initial set; The initial set is subjected to distribution analysis to obtain hierarchical density partitioning results. The density partitioning results are then subjected to reliability analysis using a confidence assessment method to obtain an intermediate set. The intermediate set is optimized to obtain the target set, and the target set is statistically processed to form a similarity distribution.

[0009] Optionally, the step of performing high-dimensional space mapping on the similarity distribution to obtain classification decision features includes: The similarity distribution vector is concatenated with the working condition code to form a joint input vector; The joint input vector is input into a lightweight feedforward neural network with one hidden layer, wherein the activation function of the hidden layer is the ReLU function, and the neuron weight matrix of the hidden layer is designed to be dynamically selected with the working condition code as the index. Specifically, multiple weight sub-matrices corresponding to different working condition codes are pre-stored, and the corresponding weight sub-matrices are loaded according to the current working condition code for this forward calculation. The feedforward neural network performs affine transformation and nonlinear activation on the joint input vector through its hidden layer to achieve a nonlinear transformation of the similarity distribution. The output of the hidden layer is mapped to the same dimension as the number of defect categories through the output layer of the feedforward neural network to generate an output vector. The output vector is subjected to Softmax normalization, and the normalized vector is used as the classification decision feature.

[0010] Optionally, the Hough transform method is used to perform edge detection processing on the texture channel image to obtain edge features. The rolling force distribution and real-time speed are used as regularization constraints to perform curve fitting processing on the edge features to generate topological features, including: The texture channel image is subjected to multi-scale spatial analysis to obtain a set of feature maps; The Hough transform is applied to the feature map set to obtain edge features, and the spatial continuity verification is performed on the edge features to obtain a continuous path structure. Using the rolling force distribution and real-time speed as regularization constraints, a piecewise fitting method is used to smooth the trajectory of the continuous path structure to obtain the geometric boundary. Geometric features are extracted from the geometric boundary, and the geometric features are structurally integrated to generate topological features that characterize the defect morphology. The geometric features include spatial location and morphological parameters.

[0011] Optionally, the step of performing a Hough transform on the feature map set to obtain edge features, and then performing spatial continuity verification on the edge features to obtain a continuous path structure, includes: The Hough transform is used to perform spatial mapping processing on the feature map set to obtain spatial distribution data. Based on the spatial distribution data, peak extraction processing is performed to obtain edge features. The edge features are grouped based on spatial proximity to obtain candidate feature groups. The candidate feature groups are then subjected to directional consistency processing to obtain optimized feature groups. The optimized feature group is connected to obtain a preliminary path segment, and the preliminary path segment is then optimized to obtain a continuous path structure.

[0012] Optionally, the step of stitching together all the surface images based on the final rolling temperature and rolling force distribution to obtain a topological image of the ultra-wide and thick plate synchronized with the process includes: Based on the final rolling temperature and rolling force distribution, the prior information of deformation is determined by a thermo-mechanical coupled finite element model; Based on the deformation prior information, matching non-rigid transformation parameters are obtained from the non-rigid transformation parameter library for image geometric correction. The non-rigid transformation parameter library contains non-rigid transformation parameters corresponding to different temperature and pressure ranges. Geometric deformation correction is performed on each surface image based on the matched non-rigid transformation parameters to generate an image sequence; The image sequence is registered and fused based on feature points to eliminate seams, resulting in a topological image of the ultra-wide and thick plate that corresponds to the actual physical size of the plate and is synchronized with the rolling process state.

[0013] Secondly, this application provides an online detection system for surface defects in ultra-wide and thick plates based on image recognition, comprising: The acquisition module is used to acquire surface images and process status parameters at multiple locations of the ultra-wide and thick plate. The process status parameters include final rolling temperature, rolling force distribution, laminar cooling intensity, and real-time speed. The processing module is used to stitch together all the surface images according to the final rolling temperature and rolling force distribution to obtain a topological image of the ultra-wide and thick plate synchronized with the process. The decomposition module is used to input the topological image to the physically guided feature decoupling module. The feature decoupling module decomposes the topological image into a texture channel image representing defects and a background channel image representing the substrate material and oxide layer through a reflectivity separation algorithm, constrained by the laminar cooling intensity. The detection module is used to perform edge detection processing on the texture channel image using the Hough transform method to obtain edge features, and to perform curve fitting processing on the edge features using the rolling force distribution and real-time speed as regularization constraints to generate topological structure features. The analysis module is used to analyze the topological features and the preset feature library using a support vector machine. The kernel function of the support vector machine is selected by dynamically adjusting the final rolling temperature and laminar cooling intensity to obtain defect detection results.

[0014] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of the image recognition-based online detection method for surface defects in ultra-wide and thick plates as described in the first aspect above.

[0015] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the online detection method for surface defects of ultra-wide and thick plates based on image recognition as described in the first aspect above.

[0016] The online surface defect detection method for ultra-wide and thick plates based on image recognition provided in this application has the following beneficial effects: First, it simultaneously acquires surface images and key process parameters to provide a comprehensive real-time data foundation for processing; then, it performs image stitching based on process parameters to compensate for geometric distortion caused by process deformation, thereby obtaining a topological image corresponding to the physical state of the plate; next, it uses laminar cooling intensity constraints to decouple features, thereby separating defect texture channels and effectively suppressing background noise; then, it uses Hough transform to detect edges and combines rolling force and speed constraints for curve fitting, making the extracted contour features more consistent with the real deformation law, thereby generating stable topological structure features; finally, it uses a support vector machine to dynamically adjust the kernel function according to process parameters for analysis, enabling the recognition model to adapt to different working conditions and output accurate detection results.

[0017] Furthermore, this application first uses a support vector machine to dynamically construct a hybrid kernel function based on process parameters and calculate the similarity distribution; then, it obtains classification decision features through high-dimensional mapping; and finally, through pattern recognition and combined with geometric information, it generates a complete result containing type, size and location. This process enables the classifier to flexibly adjust the calculation method according to specific process conditions, so as to maintain high-precision recognition and positioning capabilities when facing changes in defect features caused by process fluctuations. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic flowchart illustrating an online detection method for surface defects in ultra-wide and thick plates based on image recognition, provided in an embodiment of this application; Figure 2 A schematic diagram illustrating a specific implementation of an online detection method for surface defects in ultra-wide and thick plates based on image recognition, provided in an embodiment of this application. Figure 3A schematic diagram of the structure of an online detection system for surface defects of ultra-wide and thick plates based on image recognition, provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] In the production of ultra-wide and thick plates, reliable online detection of surface defects is crucial to ensuring product quality. However, existing detection technologies typically employ fixed procedures and cannot respond to real-time changes in key process parameters such as final rolling temperature and rolling force. These changes directly affect the surface condition and defect morphology of the plates, leading to misjudgments and missed detections when production conditions fluctuate. Furthermore, the stability and accuracy of the detection results are difficult to guarantee.

[0021] To address this, this application proposes an online surface defect detection method that deeply integrates process parameters. The core of this method lies in enabling the detection system to dynamically perceive and adapt to real-time operating conditions. Specifically, this application first synchronously acquires images and process parameters, and then performs precise image correction and feature separation based on the parameters to suppress interference and extract pure defect information. Subsequently, process parameters are introduced as physical constraints in the feature analysis stage, making the feature characterization more consistent with actual deformation patterns. Finally, a key recognition module dynamically adjusts its internal parameters according to real-time process conditions, thereby forming an adaptive recognition capability. Therefore, this scheme, by deeply embedding dynamic process information into the entire detection process, enables the system to proactively match complex and ever-changing production realities, improving the reliability and accuracy of defect detection in real industrial scenarios.

[0022] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] The core of this application is to provide an online detection method for surface defects in ultra-wide and thick plates based on image recognition. A flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes: S101. Collect surface images and process status parameters at multiple locations of the ultra-wide and thick plate. The process status parameters include final rolling temperature, rolling force distribution, laminar cooling intensity, and real-time speed.

[0024] Among them, the final rolling temperature refers to the actual temperature reached on the surface of the plate when the final rolling process is completed; the rolling force distribution refers to the magnitude and variation of the pressure applied by the rolls to different width and length areas of the plate during the rolling process. Laminar flow cooling intensity refers to the control parameters such as flow rate and pressure of water in the laminar flow cooling system that the sheet undergoes after leaving the rolling mill; real-time speed refers to the instantaneous moving speed of the sheet along the production line at the inspection station.

[0025] In step S101, an industrial camera array is deployed above a specific workstation on the production line. The multiple cameras in the array are pre-calibrated and arranged to ensure that they can simultaneously cover the entire width range of the ultra-wide and thick plate, thereby synchronously capturing local high-resolution images of multiple locations on the plate surface.

[0026] At the same time, the system is connected in real time to the main control system of the rolling line through a data interface. It automatically reads the process status parameters, accurate to the moment, corresponding to the plate segment being photographed. These parameters are measured and uploaded in real time by various sensors distributed on the production line, ensuring strict synchronization and correspondence between image information and process status information in time and space.

[0027] S102. Based on the final rolling temperature and rolling force distribution, all the surface images are stitched together to obtain a topological image of the ultra-wide and thick plate synchronized with the process.

[0028] In one specific implementation, step S102 includes: Step 1021: Based on the final rolling temperature and rolling force distribution, determine the prior information of deformation using a thermo-coupled finite element model.

[0029] The thermo-mechanical coupled finite element model refers to a dedicated computer simulation program that uses the final rolling temperature as the thermal input and the rolling force distribution as the mechanical input. It simulates and outputs the deformation state of the sheet metal by solving the coupled equations of solid mechanics and heat conduction.

[0030] It should be noted that the structure and specific implementation process of the thermo-coupled finite element model, as well as the mathematical expressions of the coupling equations, can be referred to relevant technologies. Therefore, the embodiments of this application will not elaborate on the structure and specific implementation process of the thermo-coupled finite element model, as well as the mathematical expressions of the coupling equations.

[0031] In step 1021, the collected final rolling temperature and rolling force distribution data are input into a pre-configured thermo-mechanical coupled finite element model. After the model starts calculation, it will accurately simulate the deformation process of the plate under the current thermo-mechanical conditions based on the built-in material physical properties, and output a detailed dataset containing spatial displacement vectors, i.e. deformation prior information.

[0032] Step 1022: Based on the deformation prior information, obtain matching non-rigid transformation parameters from the non-rigid transformation parameter library for image geometric correction. The non-rigid transformation parameter library contains non-rigid transformation parameters corresponding to different temperature and pressure ranges.

[0033] Among them, the non-rigid transformation parameter library refers to a lookup table that has been established in advance through a large number of experiments and simulations. The library provides various possible combinations of final rolling temperature and rolling force distribution, and each combination interval is associated with a specific set of mathematical transformation parameters. These parameters directly define how to reverse adjust the pixel coordinates of an image that has undergone corresponding deformation.

[0034] In step 1022, the obtained deformation prior information is used as the feature basis to compare and search in the non-rigid transformation parameter library. The purpose is to accurately find the set of pre-set non-rigid transformation parameters that best match the current deformation characteristics from the library, thereby providing a direct mathematical tool for subsequent correction.

[0035] Step 1023: Perform geometric deformation correction on each surface image based on the matched non-rigid transformation parameters to generate an image sequence.

[0036] In step 1023, the acquired non-rigid transformation parameters are applied to each surface image acquired in S101. The application process can be as follows: based on the mapping relationship defined by the non-rigid transformation parameters, the position of each pixel in the surface image under the ideal flat state is recalculated, and a new pixel grid is filled using an image interpolation algorithm to generate a corrected image. After this operation is performed on all images, the image sequence is obtained. It should be noted that the mapping relationship, image interpolation algorithm, etc. are technologies known in the art, and their specific implementation process will not be described in detail in this application.

[0037] Step 1024: Perform feature point-based registration and fusion on the image sequence to eliminate seams and obtain a topological image of the ultra-wide and thick plate that corresponds to the actual physical size of the plate and is synchronized with the rolling process state.

[0038] In step 1024, adjacent images in the image sequence are first automatically aligned using a feature point detection and matching algorithm to ensure they are aligned at the splicing point. Then, multi-resolution image fusion technology is used to smooth the transition of pixels in the overlapping area to eliminate the splicing boundary. Finally, a seamless, geometrically accurate topological image of the ultra-wide and thick plate corresponding to the current rolling state is output. It should be noted that the feature point detection and matching algorithm and multi-resolution image fusion technology are well-known technologies in the field, and their specific implementation process will not be described in detail in this application.

[0039] This application achieves accurate and adaptive stitching of multiple local images by using process parameters to drive physical simulation and guide image geometric correction, thereby effectively eliminating the systematic geometric errors caused by the physical deformation of the board material to the imaging. The generated topological image provides a complete board surface view with accurate spatial reference for the subsequent extraction and identification of defect features.

[0040] S103. The topology image is input to the physically guided feature decoupling module. The feature decoupling module uses the laminar cooling intensity as a constraint and decomposes the topology image into a texture channel image representing defects and a background channel image representing the substrate material and oxide layer through a reflectivity separation algorithm.

[0041] Among them, the feature decoupling module refers to a specially designed image processing unit. The core feature of this module is that its operation logic integrates the physical knowledge in the rolling process, rather than a pure mathematical transformation. It aims to separate surface information of different causes from the mixed image based on physical laws. The reflectivity separation algorithm is a core calculation method that runs inside the feature decoupling module. This algorithm is based on the principle that different surface materials, such as steel substrate, iron oxide scale, and various defects, have inherent differences in their light reflection characteristics. It then analyzes and calculates the brightness and color values ​​of the input image pixels to estimate and distinguish the components in the image that come from different materials. Texture channel images refer to the first type of images output after processing by the feature decoupling module. These images mainly retain the information on brightness and texture changes caused by surface defects such as scratches, cracks, and pits, which appear abruptly or irregularly in local areas, while suppressing uniform backgrounds as much as possible. Defects refer to iron oxide scale peeling, microcracks, etc.; the background channel image refers to the second type of image output after processing by the feature decoupling module. This image mainly contains color and brightness information that varies slowly over a wide range, contributed by the color of the metal substrate itself and the oxide layer covering its surface.

[0042] It should be noted that the embodiments of this application do not limit the specific structure of the feature decoupling module, the training method, or the specific implementation process of the reflectivity separation algorithm, and can be set accordingly according to the actual situation.

[0043] In step S103, the topology image is sent to the physically guided feature decoupling module. While receiving the topology image, the module also acquires the real-time laminar cooling intensity parameters. Then, the reflectivity separation algorithm starts working inside the module. The algorithm first dynamically adjusts the model parameters related to the thickness and uniformity of the oxide layer based on the laminar cooling intensity parameters. This is because the cooling intensity directly determines the formation state of the oxide layer, thereby affecting the reflectivity of the background. This adjustment process is the embodiment of "constrained by laminar cooling intensity".

[0044] Subsequently, guided by the adjusted physical model, the algorithm decomposes and calculates the value of each pixel in the topological image, that is, it solves the grayscale or color information of each pixel into the superposition of two contributions: one part comes from the possible defect texture, and the other part comes from the background composed of the matrix and oxide layer. Then, through iterative optimization, it finally calculates two independent values ​​for each pixel, which are respectively attributed to the defect and the background.

[0045] Finally, after the calculation is completed, the feature decoupling module recombines the values ​​of all pixels belonging to the defect part into a complete image, namely the texture channel image; at the same time, it combines the values ​​of all pixels belonging to the background part into another image, namely the background channel image. It should be noted that the reflectivity separation algorithm is a well-known technology in the field, and this application will not elaborate on it in detail.

[0046] This application introduces laminar flow cooling intensity as a key process parameter as a physical constraint to drive the feature decoupling module to intelligently separate the surface image. This effectively removes background interference caused by material and oxide layer, thereby highlighting the real defect texture information, thus providing a clean and high-quality input image for subsequent accurate edge detection and feature extraction.

[0047] S104. The texture channel image is processed by the Hough transform method to obtain edge features. The rolling force distribution and real-time speed are used as regularization constraints to perform curve fitting on the edge features to generate topological features. In one specific implementation, step S104 includes: Step 1041: Perform multi-scale spatial analysis processing on the texture channel image to obtain a set of feature maps.

[0048] Multi-scale spatial analysis processing can refer to processing the same image multiple times using filter kernels of different sizes to extract image structure information of different thicknesses. In this step, the purpose is to simultaneously capture potential defect edges of different widths from the texture channel image.

[0049] In step 1041, the texture channel image is filtered and gradients are calculated using a set of Gaussian kernels with gradually increasing standard deviations. Specifically, first, a Gaussian kernel with a small standard deviation is used to convolve the texture channel image to obtain a feature response map that preserves fine texture details. Then, the standard deviation of the Gaussian kernel is gradually increased, and the convolution and gradient calculation are repeated to obtain a feature response map that highlights the broad and narrow contours. Finally, all feature response maps generated at different scales are arranged in scale order to form a feature map set.

[0050] For example, filtering is performed using three Gaussian kernels of different scales, with standard deviations σ set to 1 pixel, 2 pixels, and 4 pixels, respectively. Then, the gradient magnitude is calculated for each filtered image to generate three feature maps corresponding to fine, medium, and coarse edge responses, respectively. These three images constitute a set of feature images.

[0051] Step 1042: Perform Hough transform on the feature map set to obtain edge features, and perform spatial continuity verification on the edge features to obtain a continuous path structure.

[0052] Among them, edge features refer to the line segments or arc segments formed by connecting pixel positions that represent abrupt changes in grayscale in the image, which are initially identified through Hough transform. These features are usually discrete and incomplete. Spatial continuity verification processing refers to the process of analyzing edge features to determine whether they should belong to the same continuous contour in space. This processing is based on the endpoint distance and directional angle between edge features. A continuous path structure refers to a spatially continuous and directionally smooth polyline or curve formed by connecting multiple edge features belonging to the same contour after spatial continuity verification processing.

[0053] Step 1042 may specifically include the following steps: Step a1: Perform spatial mapping processing on the feature map set using Hough transform to obtain spatial distribution data. Based on the spatial distribution data, perform peak extraction processing to obtain edge features.

[0054] In step a1, the Hough transform is used to process each image in the feature map set. That is, by mapping the edge points in the image space to the parameter space for cumulative voting, the straight line and curve primitives in the image are identified. Then, by finding the local peaks in the parameter space, the corresponding edge features are extracted.

[0055] For example, for the feature map set { Each image in the diagram is subjected to a Hough transform in sequence, and the parameter space is defined. The threshold for identifying the peak vote count is T. v=45, and then after performing the Hough transform, a local peak search is performed on the parameter space distribution of each image, that is, to find those points whose vote count not only exceeds 45, but also has the maximum value in its 3×3 neighborhood, and each such peak point corresponds to a line segment primitive in the image space. Subsequently, a total of 153 short line segments satisfying the conditions are extracted from the three feature maps in this way, and each line segment is represented by its starting coordinates. End point coordinates The edge feature is described by its length l and direction angle θ, and the set of these line segments constitutes the edge feature. .

[0056] Step a2: Group the edge features based on spatial proximity to obtain candidate feature groups, and perform directional consistency processing on the candidate feature groups to obtain optimized feature groups.

[0057] It should be noted that the relevant explanations of spatial proximity can be found in related technologies, and will not be elaborated here.

[0058] In step a2, edge features that are close to each other are grouped into the same candidate group based on the principle of spatial proximity. Then, directional consistency processing is performed on each candidate group, and the average direction of the edge features within the group is calculated. Abnormal edge features with excessive directional deviation are then removed to obtain the optimized group.

[0059] For example, edge features based on the principle of spatial proximity Group the data and set a proximity threshold. =10 pixels; then calculate the midpoint coordinates of all edge feature segments. And for any two line segments E m and E n The corresponding center point coordinates are ( , )and( , If the distance of its midpoint is... If they are not found in the previous group, they are grouped into the same candidate group. Then, based on this rule, the 153 edge features are initially divided into 14 candidate groups. ; Next, directional consistency processing is performed on each candidate group, and a directional deviation threshold is set. In the group Calculate the average of the orientation angles of all edge features within the area. And remove all groups that meet the criteria. The edge features are considered, and it is assumed that after this screening, 4 candidate groups are broken up due to internal orientation inconsistency. Their members are then redistributed to other groups based on spatial proximity and orientation consistency, thus ultimately forming 10 optimized groups with better orientation consistency.

[0060] Step a3: Perform connection processing on the optimized feature group to obtain preliminary path segments, and perform optimization processing on the preliminary path segments to obtain a continuous path structure.

[0061] In step a3, it is first determined whether the endpoint distances of each edge feature in the optimization group are close and whether the directional angles are compatible. Then, the edge features that meet the above conditions are connected to form a preliminary path segment. Finally, the preliminary path segment is smoothed by spline interpolation and isolated segments that are too short are removed to output a continuous path structure.

[0062] For example, set an endpoint connection distance threshold of 5 pixels and a directional angle tolerance for the optimization group. Then, iterate through all edge features within each optimization group and calculate the minimum Euclidean distance between the four endpoints of any two line segments. ,like and Then they are connected at the pair of nearest endpoints, where, This represents the maximum endpoint distance threshold allowed to connect two detected line segments in the image space, and It can be dynamically adjusted according to the plate speed, surface condition, or imaging resolution to adapt to the edge continuity characteristics under different working conditions. and These refer to the angle between the directions of the two features, which connects multiple short line segments into a longer polyline and generates four initial path segments; Then, a minimum effective length of 20 pixels is set, and cubic spline interpolation is used to smooth the discrete point sequence of each path segment. Then, the original polyline length of each path segment before smoothing is calculated, and path segments with a length less than the minimum effective length of 20 pixels are removed. Assuming that a short path segment with a length of 15 pixels is removed after this optimization, the remaining 3 path segments are further smoothed and connected end to end, and finally merged into a smooth continuous path structure with a total length of 315 pixels.

[0063] Step 1043: Using the rolling force distribution and real-time speed as regularization constraints, the continuous path structure is smoothed using a piecewise fitting method to obtain the geometric boundary.

[0064] In this context, regularization constraints refer to conditions added in mathematical optimization to ensure that the solution conforms to certain prior knowledge. In this step, the rolling force distribution and real-time speed can be used as regularization constraints to construct a constraint function. This function is used to penalize fitting results that do not conform to the physical laws of material rolling deformation during curve fitting. The piecewise fitting method refers to dividing a non-smooth path into multiple segments, performing curve fitting on each segment separately, and finally smoothly connecting all the fitted curves. The geometric boundary refers to one or more smooth closed or non-closed curves that can accurately describe the defect contour obtained by the piecewise fitting method.

[0065] It should be noted that the embodiments of this application do not specifically limit the expression of the constraint function or the expression of the piecewise fitting method.

[0066] In step 1043, the continuous path structure is first discretized into a series of ordered scattered points. Then, a piecewise cubic spline curve is used to fit these scattered points. During the fitting process, a regularization term defined by process parameters is introduced. The weight coefficient of this regularization term is dynamically calculated based on the rolling force distribution and real-time speed. Specifically, at a certain position on the plate, the regularization weight coefficient w is directly proportional to the rolling force at that position and inversely proportional to the real-time speed at that position. The formula for calculating the regularization weight coefficient is as follows: Where F is the real-time rolling force at the corresponding position, and v is the real-time velocity. and The reference rolling force and reference speed are used, and α and β are the balance coefficients of the corresponding terms. Then, in the high rolling force or low speed region, the fitted curve will be subject to stronger smoothing constraints to suppress overfitting caused by noise or discontinuities, thereby generating geometric boundaries.

[0067] For example, obtain the real-time process parameters at the location of the plate corresponding to the continuous path structure: rolling force F = 18MN, real-time speed v = 1.2m / s, and set reference process parameters. , We set α and β to 0.5, and then obtained w equal to 1.225 according to the regularization weight coefficient formula. Subsequently, when fitting the discrete point set of the piecewise cubic spline curve pair, we introduced the weight w=1.225 into the regularization term in the objective function. This makes the algorithm apply a stronger smoothness constraint on the path segment during the fitting process, and effectively suppresses the overfitting or distortion of the curve caused by image noise or local discontinuities. This generates a smooth curve that better conforms to the physical trend of material deformation under high rolling force and low speed conditions, i.e., the geometric boundary.

[0068] Step 1044: Extract geometric features from the geometric boundary, perform structured integration of the geometric features, and generate topological features that characterize the defect morphology, wherein the geometric features include spatial location and morphological parameters.

[0069] Among them, geometric features refer to the quantified parameters extracted from the geometric boundary. These features may include the two-dimensional coordinates of key points, the length and direction angle of the line segments connecting adjacent key points, and the area of ​​the closed region enclosed by the boundary. Topological features refer to the data set obtained after the geometric features are structured and organized. This set not only contains geometric parameters, but also clearly describes the connection and hierarchical relationships between the boundary line segments.

[0070] In step 1044, all key points on the geometric boundary are traversed and their coordinates are recorded. Then, the length and direction angle of the line segment connecting adjacent key points are calculated. Next, the area enclosed by the closed boundary is calculated. Then, the spatial relationship between the boundaries is integrated, that is, a connection table between the boundary endpoints is established to describe the topological connectivity. Then, for nested closed boundaries, a parent-child hierarchical relationship is established. Finally, all geometric parameters and topological relationship data are organized into a structured data object, that is, topological structure features.

[0071] For example, firstly, sampling is performed on the boundary curve at fixed arc length intervals, combined with the detection of curvature extreme points, to determine a total of 85 key points, and the coordinate sequence formed by the coordinates of all key points is recorded. Then, the line segments formed by adjacent key points are calculated, resulting in a total of 84 line segments. The length and orientation angle of each line segment are calculated. Since this crack profile is a non-closed curve, the area it encloses is approximately 0. Next, a topological connection table is created, such as using an 84×2 matrix to record the predecessor and successor key point indices of each line segment to describe the connection order of the boundary points; then all geometric parameters, namely: key point coordinate set, line segment length set, direction angle set, area, and topological connection relationship, are encapsulated into a hierarchical structural data object as the topological structural feature of this crack, and its type is marked as "open curve".

[0072] This application ensures the complete detection of defect edges through multi-scale analysis and innovatively transforms real-time process parameters into physical constraints for curve fitting, making the extracted defect contour features more consistent with the material rolling deformation law. This not only effectively suppresses the fitting deviation caused by image noise and interference, but also lays a solid foundation for the accurate identification of subsequent defect types.

[0073] S105. The topological features and the preset feature library are analyzed using a support vector machine. The kernel function of the support vector machine is selected by dynamically adjusting the final rolling temperature and laminar cooling intensity to obtain the defect detection results.

[0074] The defect detection results can include defect location information and defect type.

[0075] In one specific implementation, such as Figure 2 As shown, step S105 includes: Step 1051: Based on the final rolling temperature and laminar cooling intensity, query the preset process condition mapping table to determine the condition code to which the current state belongs.

[0076] The process condition mapping table is a predefined data lookup table that uses different combinations of final rolling temperature and laminar cooling intensity as input keys to assign a unique identifier to each typical process condition combination.

[0077] In step 1051, based on the real-time collected final rolling temperature and laminar cooling intensity, the process condition mapping table is queried. This table defines unique codes corresponding to different temperature and cooling intensity ranges. The currently collected specific values ​​are compared with the ranges in the table to find matching entries, and their corresponding codes are used as the condition codes to which the current state belongs.

[0078] For example, if the current final rolling temperature is 920°C and the laminar cooling intensity is "medium", and the preset process condition mapping table is queried, there is a record in the table that specifies that when the temperature is between 900°C and 950°C and the cooling intensity is "medium", the corresponding condition code is "G1". The current parameters meet this condition, so the current condition code is determined to be G1.

[0079] Step 1052: Using a support vector machine, based on the kernel function configuration scheme corresponding to the working condition code, the linear kernel and radial basis kernel of the support vector machine are dynamically combined, and the parameters of each kernel are set to construct a hybrid kernel function. Based on the hybrid kernel function, the similarity between the topological structure features and the feature templates in the preset feature library is calculated to obtain the similarity distribution.

[0080] Among them, the kernel function configuration scheme refers to a set of parameter settings that are pre-associated with each working condition code. The scheme specifies the mixing weights and core parameters of each component used to construct the hybrid kernel function of the support vector machine under a specific working condition. Hybrid kernel functions are new kernel functions formed by combining linear kernels and radial basis kernels, which are used to calculate the similarity between features more flexibly.

[0081] It should be noted that the specific expressions of the hybrid kernel function, linear kernel, and radial basis kernel in the embodiments of this application are not limited, and can be set accordingly according to the actual situation.

[0082] Step 1052 may specifically include the following steps: Step b1: Based on the hybrid kernel function, calculate the high-dimensional spatial similarity between the topological structure features and the feature templates of various defects in the preset feature library to obtain an initial set.

[0083] In step b1, based on the queried working condition code, the corresponding kernel function configuration scheme is obtained. This scheme specifies the mixing ratio of the linear kernel and the radial basis function kernel, as well as the key parameters of the radial basis function kernel. Then, according to the ratio parameters in the scheme, the linear kernel and the radial basis function kernel are added together according to their weights to construct a mixed kernel function. Subsequently, this mixed kernel function is used to calculate the similarity between the topological features to be detected and the standard templates of each type of defect in the feature library, thereby obtaining a set of original values ​​and forming an initial set.

[0084] For example, based on the working condition code G1, its kernel function configuration scheme is found to be: the linear kernel weight accounts for 70%, and the radial basis kernel parameter is 0.01. Therefore, the hybrid kernel function is composed of 70% linear kernel and 30% radial basis kernel. Using this function, the similarity between the feature to be tested and the "crack" template in the feature library is calculated to be 0.82, and the similarity between it and the "depression" template is 0.45, thus obtaining the initial set {0.82, 0.45}.

[0085] Step b2: Perform distribution analysis on the initial set to obtain hierarchical density partitioning results, and perform reliability analysis on the density partitioning results using a confidence assessment method to obtain an intermediate set.

[0086] Density partitioning results refer to the multiple levels or intervals divided according to the density of the clustering of the initial similarity values ​​on the number axis after statistical analysis.

[0087] In step b2, firstly, a clustering method is used to divide the initial similarity into several levels based on the numerical value to obtain a hierarchical density partitioning result; then, the numerical stability of each level is evaluated, that is, the dispersion of the values ​​within each level is calculated, such as the standard deviation, and the smaller the dispersion, the higher the confidence of the level is considered; then, the calculated level confidence score is used to weight and correct each initial similarity value belonging to the level, and the corrected values ​​constitute the intermediate set.

[0088] For example, we set 0.65 as the dividing line and divide the similarity into a high level (>0.65) and a low level (≤0.65). The high-level values ​​have a high confidence level, so we assume their weight is 0.9. The low-level values ​​are relatively dispersed and have a low confidence level, so we assume their weight is 0.6. Then we adjust the original values ​​with the weights, such as 0.82 × 0.9 = 0.738 and 0.45 × 0.6 = 0.27, thus obtaining the intermediate set {0.738, 0.27}.

[0089] Step b3: Optimize the intermediate set to obtain the target set, and perform statistical processing on the target set to form a similarity distribution.

[0090] In step b3, a threshold is set, and the similarity values ​​in the intermediate set below this threshold are set to zero to filter out weak or meaningless signals, thus obtaining the target set. Finally, all the values ​​in this set are normalized so that their sum of squares is 1, thereby forming a normalized vector that can be used for subsequent comparisons, i.e., the similarity distribution.

[0091] For example, if the filtering threshold is set to 0.1, since all values ​​in the intermediate set are greater than 0.1, the target set remains {0.738, 0.27}. This set is then normalized: first, the sum of squares is calculated to be 0.738. 2 +0.27 2 =0.617, and take the square root to get 0.786; then divide each value by 0.786 to get the final similarity distribution of approximately [0.94, 0.34], which means that the similarity with "crack" is 0.94 and the similarity with "dent" is 0.34.

[0092] Step 1053: Perform high-dimensional space mapping processing on the similarity distribution to obtain classification decision features.

[0093] Step 1053 may specifically include the following steps: Step c1: Concatenate the similarity distribution vector with the working condition code to form a joint input vector.

[0094] For example, the working condition code G1 is converted into a three-bit binary code [1, 0, 0], and the similarity distribution [0.94, 0.34] is concatenated with it to obtain the joint input vector. =[0.94, 0.34, 1, 0, 0].

[0095] Step c2: Input the joint input vector into a lightweight feedforward neural network with one hidden layer, wherein the activation function of the hidden layer is the ReLU function, and the neuron weight matrix of the hidden layer is designed to be dynamically selected with the working condition code as the index. Specifically, multiple weight sub-matrices corresponding to different working condition codes are pre-stored, and the corresponding weight sub-matrices are loaded according to the current working condition code for this forward calculation.

[0096] The lightweight feedforward neural network refers to a simple multilayer perceptron, specifically a network with one hidden layer used for nonlinear feature transformation in this step. The hidden layer neuron weight matrix refers to a model parameter sharing mechanism, which is implemented by pre-training and storing multiple sets of different fully connected layer weight matrices, each corresponding to a working condition code. Then, during the forward computation, the corresponding weight matrix is ​​dynamically loaded according to the current working condition code for computation.

[0097] For example, the joint input vector The input is fed into a preset lightweight feedforward neural network, and G1 is encoded according to the current operating condition. The system dynamically selects and loads the hidden layer weight matrix W corresponding to G1 from three pre-stored weight sub-matrices, namely G1, G2, and G3 respectively. h For this calculation, we assume W... h It is a 5×4 matrix.

[0098] Step c3: Perform affine transformation and nonlinear activation on the joint input vector through the hidden layer of the feedforward neural network to achieve nonlinear transformation of the similarity distribution.

[0099] For example, using the hidden layer weight matrix W h and the corresponding hidden layer bias vector The affine transformation and nonlinear activation are performed on the joint input vector v. The calculation process is as follows: First, use... Perform a linear transformation to a 4-dimensional intermediate vector z, then apply the ReLU activation function to each element of z. This is used to generate the final output vector h of the hidden layer, assuming that after calculation, h = [0.6, 1.3, 0.1, 0.9].

[0100] Step c4: The output of the hidden layer is mapped to the same dimension as the number of defect categories through the output layer of the feedforward neural network to generate an output vector.

[0101] For example, the output vector h of the hidden layer is input into the output layer of the neural network, and the output layer has a fixed weight matrix W. o Its dimensions are 4×2 and its bias vector is b. o The dimension of h is 2, and then the 4-dimensional h is mapped to a 2-dimensional space with the same number of defect categories. The calculation process is as follows: Where o represents the output vector, assuming the calculated output vector is o=[2.2, 0.3].

[0102] Step c5: Perform Softmax normalization on the output vector and use the normalized vector as the classification decision feature.

[0103] Softmax normalization refers to transforming the output vector of a neural network into a probability distribution such that the sum of all output values ​​is 1, and each value represents the probability of belonging to the corresponding category.

[0104] For example, the output layer vector o is Softmax normalized and transformed into a probability distribution, i.e., a classification decision feature; then the probability of each class is calculated: first, the exponent term is calculated. , Then calculate the sum 9.03 + 0.741 = 9.771. Finally, divide each exponent by the sum to get the normalized probabilities: P1 = 9.03 / 9.771 ≈ 0.924, P2 = 0.741 / 9.771 ≈ 0.076. Therefore, the final classification decision feature is the probability vector p ≈ [0.924, 0.076].

[0105] Step 1054: Use a classification mechanism to perform pattern recognition processing on the classification decision features to obtain the defect type. Based on the defect type, and combined with the defect geometric information parsed from the topological features, determine the size and location coordinates of the defect.

[0106] In this step, the classification mechanism specifically refers to a decision rule, which selects the category corresponding to the highest probability value among the classification decision features as the recognition result. This embodiment does not limit the specific content of the mechanism and can be set according to the actual situation. Defect geometric information refers to the data describing the size and location of defects directly parsed from the topological features.

[0107] In step 1054, a classification mechanism is used to find the probability value with the largest value in the probability vector of the classification decision feature, and the corresponding defect category is determined as the final defect type. At the same time, the pre-calculated defect geometric information, such as the length, width, area and center point coordinates of the circumscribed rectangle, is extracted from the initially input topological feature data to determine the size and specific location of the defect.

[0108] For example, in the classification decision feature [0.924, 0.076], the category corresponding to the maximum probability value of 0.924 is "crack", so the defect type is determined to be crack. At the same time, the length of the defect is 15mm and the width is 0.5mm, and the coordinates of its center point are (2050mm, 350mm) read from the topological structure feature.

[0109] Step 1055: The defect type, size, and location coordinates are fused to generate a defect detection result.

[0110] For example, by combining the three pieces of information—"crack" type, size "15mm x 0.5mm", and location "(2050, 350)"—the final defect detection result can be expressed as: A crack with a length of 15 mm was identified at coordinates (2050, 350).

[0111] Figure 3 This is a schematic diagram illustrating a specific implementation of an online surface defect detection system for ultra-wide and thick plates based on image recognition, as provided in this application. (Refer to...) Figure 3 The system may include: The acquisition module 31 is used to acquire surface images and process status parameters at multiple locations of the ultra-wide and thick plate. The process status parameters include final rolling temperature, rolling force distribution, laminar cooling intensity, and real-time speed.

[0112] The processing module 32 is used to stitch together all the surface images according to the final rolling temperature and rolling force distribution to obtain a topological image of the ultra-wide and thick plate synchronized with the process.

[0113] The decomposition module 33 is used to input the topological image to the physically guided feature decoupling module. The feature decoupling module uses the laminar cooling intensity as a constraint and decomposes the topological image into a texture channel image representing defects and a background channel image representing the substrate material and oxide layer through a reflectivity separation algorithm.

[0114] The detection module 34 is used to perform edge detection processing on the texture channel image using the Hough transform method to obtain edge features, and to perform curve fitting processing on the edge features using the rolling force distribution and real-time speed as regularization constraints to generate topological structure features.

[0115] Analysis module 35 is used to analyze the topological features and the preset feature library using a support vector machine. The kernel function of the support vector machine is selected by dynamically adjusting the final rolling temperature and laminar cooling intensity to obtain defect detection results.

[0116] The image recognition-based online detection system for ultra-wide and thick plate surface defects in this application is used to implement the aforementioned image recognition-based online detection method for ultra-wide and thick plate surface defects. Therefore, the specific implementation of the image recognition-based online detection system for ultra-wide and thick plate surface defects can be found in the embodiment section of the image recognition-based online detection method for ultra-wide and thick plate surface defects described above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0117] like Figure 4As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of the above-described image recognition-based online detection method for surface defects of ultra-wide and thick plates.

[0118] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described image recognition-based online detection methods for surface defects in ultra-wide and thick plates.

[0119] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0120] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the online detection method for ultra-wide and thick plate surface defects based on image recognition.

[0121] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0122] The above provides a detailed description of the online detection method and system for ultra-wide and thick plate surface defects based on image recognition provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. An online detection method for surface defects in ultra-wide and thick plates based on image recognition, characterized in that, include: Surface images and process parameters of ultra-wide and thick plates are collected from multiple locations. The process parameters include final rolling temperature, rolling force distribution, laminar cooling intensity, and real-time speed. Based on the final rolling temperature and rolling force distribution, all the surface images are stitched together to obtain a topological image of the ultra-wide and thick plate synchronized with the process. The topological image is input to a physically guided feature decoupling module. The feature decoupling module uses the laminar cooling intensity as a constraint and a reflectivity separation algorithm to decompose the topological image into a texture channel image representing defects and a background channel image representing the substrate material and oxide layer. The texture channel image is processed by Hough transform to obtain edge features. The rolling force distribution and real-time speed are used as regularization constraints to perform curve fitting on the edge features to generate topological features. Support vector machines (SVMs) are used to analyze the topological features and a preset feature library. The kernel function of the SVM is selected by dynamically adjusting the final rolling temperature and laminar cooling intensity. The resulting defect detection results include: Based on the final rolling temperature and laminar cooling intensity, a preset process condition mapping table is consulted to determine the condition code to which the current state belongs. Using a support vector machine, the linear kernel and radial basis kernel of the support vector machine are dynamically combined according to the kernel function configuration scheme corresponding to the working condition code, and the parameters of each kernel are set to construct a hybrid kernel function. Based on the hybrid kernel function, the similarity between the topological structure features and the feature templates in the preset feature library is calculated to obtain the similarity distribution. The similarity distribution is subjected to high-dimensional space mapping to obtain classification decision features; A classification mechanism is used to perform pattern recognition processing on the classification decision features to obtain the defect type. Based on the defect type, combined with the defect geometric information parsed from the topological structure features, the size and location coordinates of the defect are determined. The defect type, size, and location coordinates are fused together to generate a defect detection result.

2. The method according to claim 1, characterized in that, The step of calculating the similarity between the topological features and feature templates in a preset feature library based on the hybrid kernel function to obtain a similarity distribution includes: Based on the hybrid kernel function, the high-dimensional similarity between the topological structure features and the feature templates of various defects in the preset feature library is calculated to obtain an initial set; The initial set is subjected to distribution analysis to obtain hierarchical density partitioning results. The density partitioning results are then subjected to reliability analysis using a confidence assessment method to obtain an intermediate set. The intermediate set is optimized to obtain the target set, and the target set is statistically processed to form a similarity distribution.

3. The method according to claim 1, characterized in that, The high-dimensional space mapping process performed on the similarity distribution to obtain classification decision features includes: The similarity distribution vector is concatenated with the working condition code to form a joint input vector; The joint input vector is input into a lightweight feedforward neural network with one hidden layer, wherein the activation function of the hidden layer is the ReLU function, and the neuron weight matrix of the hidden layer is designed to be dynamically selected with the working condition code as the index. Specifically, multiple weight sub-matrices corresponding to different working condition codes are pre-stored, and the corresponding weight sub-matrices are loaded according to the current working condition code for this forward calculation. The feedforward neural network performs affine transformation and nonlinear activation on the joint input vector through its hidden layer to achieve a nonlinear transformation of the similarity distribution. The output of the hidden layer is mapped to the same dimension as the number of defect categories through the output layer of the feedforward neural network to generate an output vector. The output vector is subjected to Softmax normalization, and the normalized vector is used as the classification decision feature.

4. The method according to claim 1, characterized in that, The Hough transform method is used to perform edge detection processing on the texture channel image to obtain edge features. The rolling force distribution and real-time speed are used as regularization constraints to perform curve fitting processing on the edge features to generate topological structure features, including: Multi-scale spatial analysis is performed on the texture channel image to obtain a set of feature maps; The Hough transform is applied to the feature map set to obtain edge features, and the spatial continuity verification is performed on the edge features to obtain a continuous path structure. Using the rolling force distribution and real-time speed as regularization constraints, a piecewise fitting method is used to smooth the trajectory of the continuous path structure to obtain the geometric boundary. Geometric features are extracted from the geometric boundary, and the geometric features are structurally integrated to generate topological features that characterize the defect morphology. The geometric features include spatial location and morphological parameters.

5. The method according to claim 4, characterized in that, The process of performing a Hough transform on the feature map set to obtain edge features, and then performing spatial continuity verification on the edge features to obtain a continuous path structure, includes: The Hough transform is used to perform spatial mapping processing on the feature map set to obtain spatial distribution data. Based on the spatial distribution data, peak extraction processing is performed to obtain edge features. The edge features are grouped based on spatial proximity to obtain candidate feature groups. The candidate feature groups are then subjected to directional consistency processing to obtain optimized feature groups. The optimized feature group is connected to obtain a preliminary path segment, and the preliminary path segment is then optimized to obtain a continuous path structure.

6. The method according to claim 1, characterized in that, The process of stitching together all the surface images based on the final rolling temperature and rolling force distribution to obtain a topological image of the ultra-wide and thick plate synchronized with the process includes: Based on the final rolling temperature and rolling force distribution, the prior information of deformation is determined by a thermo-mechanical coupled finite element model; Based on the deformation prior information, matching non-rigid transformation parameters are obtained from the non-rigid transformation parameter library for image geometric correction. The non-rigid transformation parameter library contains non-rigid transformation parameters corresponding to different temperature and pressure ranges. Geometric deformation correction is performed on each surface image based on the matched non-rigid transformation parameters to generate an image sequence; The image sequence is registered and fused based on feature points to eliminate seams, resulting in a topological image of the ultra-wide and thick plate that corresponds to the actual physical size of the plate and is synchronized with the rolling process state.

7. An online detection system for surface defects of ultra-wide and thick plates based on image recognition, used to execute the online detection method for surface defects of ultra-wide and thick plates based on image recognition as described in any one of claims 1 to 6, characterized in that, include: The acquisition module is used to acquire surface images and process status parameters at multiple locations of the ultra-wide and thick plate. The process status parameters include final rolling temperature, rolling force distribution, laminar cooling intensity, and real-time speed. The processing module is used to stitch together all the surface images according to the final rolling temperature and rolling force distribution to obtain a topological image of the ultra-wide and thick plate synchronized with the process. The feature decoupling module is used to input the topology image into the physically guided feature decoupling module. The feature decoupling module uses the laminar cooling intensity as a constraint and decomposes the topology image into a texture channel image representing defects and a background channel image representing the substrate material and oxide layer through a reflectivity separation algorithm. The detection module is used to perform edge detection processing on the texture channel image using the Hough transform method to obtain edge features, and to perform curve fitting processing on the edge features using the rolling force distribution and real-time speed as regularization constraints to generate topological structure features. The analysis module is used to analyze the topological features and the preset feature library using a support vector machine. The kernel function of the support vector machine is selected by dynamically adjusting the final rolling temperature and laminar cooling intensity to obtain defect detection results.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the online detection method for surface defects of ultra-wide and thick plates based on image recognition as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the online detection method for surface defects of ultra-wide and thick plates based on image recognition as described in any one of claims 1 to 6.

Images