Metal product industry AI corpus processing method and system based on cloud platform

Abstract

The invention discloses a metal product industry AI corpus processing method and system based on a cloud platform, and relates to the technical field of corpus processing, and the method comprises the steps: carrying out the spangle center extraction of a gray image of a target galvanized steel strip, and obtaining a deduplicated spangle center point; determining a spangle center to which a pixel coordinate in the grayscale image belongs based on the deduplicated spangle center point; carrying out distribution analysis on the stripe candidate region corresponding to the grayscale image according to the spangle center to obtain a distribution result; performing virtual resampling analysis on the stripe candidate region, and determining sampling phase sensitivity measurement; according to the sampling phase sensitivity measurement and the distribution result, a defect label corresponding to the grayscale image is updated, and an updated label is obtained; according to the invention, the abnormal fringes amplified under a specific sampling condition can be effectively identified.

Description

Technical Field

[0001] This invention relates to the field of corpus processing technology, and in particular to an AI corpus processing method and system for the metal products industry based on a cloud platform. Background Technology

[0002] In the metal products industry, continuous hot-dip galvanizing production lines are widely used for anti-corrosion treatment of steel strip surfaces. These production lines typically include multiple processes such as pickling, fluxing, zinc pot, cooling or quenching, and online visual inspection. The steel strip runs at high speed and continuously on the production line in coil form, and the galvanized layer naturally forms a zinc flower texture with a crystalline morphology during solidification. The online visual inspection station usually uses linear or area array cameras to perform high-speed imaging of the galvanized steel strip surface and uploads the acquired surface images to a cloud platform for defect detection analysis and subsequent artificial intelligence corpus construction. With the development of cloud platforms and intelligent manufacturing, more and more galvanized steel strip surface images are being centrally stored and used as a data foundation for defect recognition algorithm training, quality statistical analysis, and process optimization decisions.

[0003] In actual production, the zinc flower crystal texture on the galvanized surface has a relatively complex spatial distribution. The online imaging process usually uses a fixed sampling interval and a fixed scanning trajectory to perform discrete sampling on the high-speed running steel strip. Under certain conditions, it is easy to generate strip-like abnormal textures at the image level. These abnormal textures are highly similar to real stripe defects in appearance, but there are no corresponding material defects on the steel strip.

[0004] Existing technologies rely solely on conventional grayscale change analysis, texture detection, or rule-based judgment methods to process images, which often fails to accurately distinguish between the aforementioned abnormal textures and real defects. This results in such abnormal textures being mislabeled as real defects and written into cloud corpora. As the corpus continues to expand, such mislabeled data will accumulate, thereby affecting the training effect of subsequent corpus-based defect recognition algorithms and reducing the reliability of detection results. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies that make it difficult to accurately distinguish between the aforementioned abnormal textures and real defects, and to propose a cloud-based AI corpus processing method and system for the metal products industry.

[0006] To address the problems existing in the prior art, the present invention adopts the following technical solution: AI-based language processing methods for the metal products industry based on cloud platforms include: The center of the zinc flower is extracted from the grayscale image of the target galvanized steel strip to obtain the center point of the deduplicated zinc flower; the center of the zinc flower to which the pixel coordinates in the grayscale image belong is determined based on the center point of the deduplicated zinc flower. Based on the zinc flower center, the distribution of the candidate stripe regions corresponding to the grayscale image is analyzed to obtain the distribution results; virtual resampling analysis is performed on the candidate stripe regions to determine the sampling phase sensitivity measure. The defect labels corresponding to the grayscale images are updated based on the sampling phase sensitivity measurement and distribution results to obtain updated labels; A corpus of defective AI data generated from updated labels and grayscale images of target galvanized steel strips.

[0007] Preferably, the specific steps for obtaining the center point of the de-zinc spangle are as follows: Obtain the grayscale image of the target galvanized steel strip; A smoothed image is obtained by smoothing a grayscale image; Based on the horizontal and vertical gradients of the smoothed image, gradient feature analysis is performed on the smoothed image to obtain several candidate points for the center of zinc flower. Spatial aggregation of all candidate zinc spangle centers yields several deduplicated zinc spangle centers.

[0008] Preferably, the specific steps for determining the center of the zinc flower are as follows: Obtain the pixel coordinates of each pixel in a grayscale image; Determine the Euclidean distance between the pixel coordinates and the center points of all deduplicated zinc spangles; The center point of the deduplicated zinc spangle corresponding to the smallest Euclidean distance is determined as the center of the zinc spangle to which the pixel coordinate belongs.

[0009] Preferably, the specific steps to obtain the distribution results are as follows: One-dimensional difference processing is performed on the pixel gray levels of a grayscale image to obtain a gray level change sequence; The grayscale image is divided into regions based on the grayscale change sequence to obtain several stripe candidate regions; Based on the zinc flower centers corresponding to all pixel coordinates within the stripe candidate region, determine the subset of zinc flower centers of the stripe candidate region. Determine the azimuth angle between the pixel coordinates and the center of its corresponding zinc flower; The azimuth distribution of the stripe candidate region is determined based on the azimuth angles corresponding to the coordinates of all pixels within the stripe candidate region. The distribution of the stripe candidate region was analyzed by azimuth angle distribution, and the distribution results were obtained.

[0010] Preferably, the specific steps for determining the sampling phase sensitivity metric are as follows: Determine the target location of the stripe candidate region in the grayscale image; Local resampling is performed on the target location to obtain several virtual resampled image patches; A one-dimensional projection of the pixel grayscale values ​​of the virtual resampled image patch yields a stripe intensity sequence. The variance of the stripe intensity sequence is used as the stripe significance value; The sampling phase sensitivity measure of the stripe candidate region is determined based on the stripe saliency values ​​corresponding to all virtual resampled image patches.

[0011] Preferably, the specific steps to obtain the updated tag are as follows: The candidate regions of the stripes are grouped according to the sampling phase sensitivity metric to obtain the grouping results; Based on the grouping and distribution results, the candidate regions of the stripes are classified, and the classification results are obtained. Obtain the defect labels corresponding to the grayscale images; The defect labels corresponding to the grayscale images are updated based on the category determination results to obtain updated labels.

[0012] Preferably, the specific steps to obtain the category determination result are as follows: If the grouping result of the stripe candidate region is a sensitive group and the distribution result is that the azimuth distribution satisfies the distribution characteristic of the azimuth angle concentrated in a single main direction, then the category determination result of the stripe candidate region is set as the zinc flower sampling aliasing pseudo-defect stripe region. If the grouping result of the stripe candidate region is a Class II sensitive group, and the distribution result is an azimuth distribution whose distribution characteristics do not satisfy the requirement that the azimuth angle is concentrated in a single main direction, then the category determination result of the stripe candidate region is set as a real stripe defect region.

[0013] Preferably, the specific steps for generating the defective AI corpus set are as follows: By updating the labels, all stripe candidate regions are classified to obtain the first stripe region set, the second stripe region set, and the third stripe region set; Obtain the coil index of the target galvanized steel strip; Based on the second stripe region set, grayscale images, and roll material index, a set of real defect corpora is constructed; The set of real defect corpora is used as the set of defect AI corpora for the target galvanized steel strip.

[0014] To address the aforementioned problems, this invention also provides a cloud-based AI corpus processing system for the metal products industry, the system comprising: The zinc spangle center determination module is used to extract the zinc spangle center from the grayscale image of the target galvanized steel strip and obtain the deduplicated zinc spangle center point; based on the deduplicated zinc spangle center point, the zinc spangle center to which the pixel coordinates in the grayscale image belong is determined. The stripe feature analysis module is used to perform distribution analysis on the stripe candidate regions corresponding to the grayscale image based on the center of the zinc flower, and obtain the distribution results; it also performs virtual resampling analysis on the stripe candidate regions to determine the sampling phase sensitivity measure. The defect label update module is used to update the defect labels corresponding to the grayscale image based on the sampling phase sensitivity measurement and distribution results, and obtain the updated labels. The AI ​​corpus generation module is used to generate a defect AI corpus set of target galvanized steel strips based on updated labels and grayscale images.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention introduces a processing method at the grayscale image level to extract the center of the zinc flower and determine the center of the zinc flower to which the pixel belongs. This makes the analysis of stripe candidate regions no longer limited to traditional brightness changes or appearance morphology judgments. Instead, it establishes a spatial correspondence between image pixels and the zinc flower structure on the galvanized surface. It performs statistical analysis on the directional distribution of pixels within the stripe candidate region relative to the center of the zinc flower, so that the overall pattern of the stripes in spatial direction can be effectively depicted. This provides a reliable structural basis for distinguishing abnormal stripes introduced by the imaging process from real material defects, and improves the rationality and stability of stripe judgment.

[0016] 2. This invention constructs a sampling phase sensitivity metric by performing virtual resampling analysis on the candidate stripe region, enabling the quantification of the significant changes in stripes under different sampling starting conditions. Starting from the imaging sampling mechanism, it reveals the response law of abnormal stripes to changes in sampling conditions, thereby avoiding static judgment of stripes based solely on a single imaging result. By combining the sampling phase sensitivity metric with the directional distribution analysis results, it can effectively identify abnormal stripes that are magnified under specific sampling conditions and distinguish them from real defects that are stable in imaging, reducing the probability of abnormal stripes being mistaken for real defects.

[0017] 3. After determining the stripe category, this invention updates the original defect labels of the grayscale image and constructs a defect corpus based on the updated labels. This ensures that the final corpus contains only defect information that is consistent with the actual surface condition of the target galvanized steel strip. By removing data corresponding to abnormal stripes during the corpus construction stage, the problem of erroneous corpus entering the cloud platform and being continuously accumulated is avoided, effectively improving the purity and credibility of the defect corpus. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1This is a flowchart illustrating a cloud-based AI corpus processing method for the metal products industry, provided as an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] This embodiment provides a cloud-based AI corpus processing method for the metal products industry. (See [link to relevant documentation]). Figure 1 Specifically, including: The center of the zinc flower is extracted from the grayscale image of the target galvanized steel strip to obtain the center point of the deduplicated zinc flower; the center of the zinc flower to which the pixel coordinates in the grayscale image belong is determined based on the center point of the deduplicated zinc flower. In an embodiment of the present invention, the specific steps for obtaining the center point of the de-zinc spangle are as follows: Obtain the grayscale image of the target galvanized steel strip; First, the operating status of the target galvanized steel strip is clearly defined, ensuring it is in a continuous hot-dip galvanizing production process, has completed the galvanizing process, and is stably entering the surface quality inspection stage. The steel strip runs continuously at a uniform speed in coil form, with the operating speed matching the subsequent inspection steps. Then, an online appearance inspection device is deployed. This device includes an industrial camera, a stable light source, and an image transmission component. The stable light source uses a uniform surface light source, and the light source intensity is adjusted to avoid excessive reflection or insufficient light on the galvanized surface. The angle between the light source and the steel strip surface is set to highlight the differences between zinc patterns and potential defects. The online appearance inspection device is activated, triggering the industrial camera to capture images according to a preset acquisition frequency. The acquisition frequency matches the steel strip's running speed, ensuring that adjacent frames are not missed or excessively overlapped along the length of the steel strip. The industrial camera converts the captured color images into grayscale images. The conversion process is achieved by calculating the weighted average of the red, green, and blue components of each pixel. The weighting coefficient is determined based on the reflectivity of the galvanized steel strip surface, ensuring that the grayscale value accurately reflects the intensity of reflection or scattering of incident light by the corresponding galvanized surface. Finally, a grayscale image representing the brightness distribution of the target galvanized steel strip surface is obtained.

[0021] The target galvanized steel strip refers to a steel strip product that has completed the galvanizing process and entered the surface quality inspection stage in the continuous hot-dip galvanizing production process. The steel strip runs continuously in the form of coils, and its surface is covered with a galvanized layer formed by the solidification of molten zinc. During the solidification process, the galvanized layer forms zinc flower texture with crystalline structure characteristics.

[0022] A smoothed image is obtained by smoothing a grayscale image; First, the acquired grayscale image is preprocessed to remove potentially invalid pixels at edges. Invalid pixels are defined as those whose grayscale values ​​exceed the normal brightness distribution range of a galvanized surface. Then, a Gaussian smoothing algorithm is used to smooth the grayscale image. First, the size of the Gaussian convolution kernel is determined based on the noise characteristics of the grayscale image. The kernel size is chosen to be an odd number, specifically determined by statistically analyzing the fluctuation range of pixel grayscale values ​​in the image. A larger fluctuation range requires a larger kernel size, and vice versa. Next, the values ​​of each element of the Gaussian convolution kernel are calculated using a Gaussian function. The standard deviation of the Gaussian function is determined based on the noise characteristics of the grayscale image. The fineness of the zinc flower texture is determined to ensure that the standard deviation can preserve the crystal structure characteristics of the zinc flower while removing noise. Then, a Gaussian convolution kernel is convolved with the grayscale image. During the operation, each pixel in the grayscale image is taken as the center, and the neighboring pixels corresponding to the size of the convolution kernel are taken. The grayscale value of the neighboring pixels is multiplied by the element value of the corresponding position of the convolution kernel, and then all the product results are summed to obtain the smoothed grayscale value of the center pixel. For pixels at the edge of the grayscale image, a mirror filling method is used to supplement the neighboring pixels to ensure that the smoothing effect of the edge pixels is consistent with that of the pixels inside the image. After completing the convolution operation pixel by pixel, a smoothed image is obtained.

[0023] A grayscale image refers to two-dimensional image data acquired by online appearance inspection equipment to characterize the brightness distribution of the target galvanized steel strip surface. The grayscale value of each pixel in the image reflects the intensity of reflection or scattering of incident light by the galvanized surface at the corresponding location. A smooth image refers to a grayscale image obtained after smoothing, and its brightness changes mainly reflect the zinc flower crystal structure and macroscopic texture characteristics of the galvanized surface.

[0024] Based on the horizontal and vertical gradients of the smoothed image, gradient feature analysis is performed on the smoothed image to obtain several candidate points for the center of zinc flower. First, the zinc flower crystal regions in the smooth image are traversed and identified. The number of pixels independently occupied by each zinc flower crystal is counted, and the arithmetic mean of the number of pixels in all zinc flower crystals is calculated to obtain the average pixel size of the zinc flower. Based on this average pixel size, an odd number suitable for the zinc flower is selected as the specification of the convolution kernel used for gradient calculation, ensuring that the convolution kernel can completely cover the core area of ​​a single zinc flower crystal without crossing the boundary of adjacent zinc flower crystals. Then, based on the basic principle of gray-level difference detection, the convolution kernel elements are designed. The elements of the horizontal convolution kernel are set according to the rule of left-right symmetrical distribution and gradient increase from the center to both sides. The left element is negative, the right element is positive, and the middle element is zero, which is used to highlight the gray-level difference of pixels in the horizontal direction. The elements of the vertical convolution kernel are symmetrically distributed vertically and gradient increase from the center to both ends. The gradient is set to increase in degree, with negative values ​​for the upper elements, positive values ​​for the lower elements, and zero values ​​for the middle elements. This is used to highlight the grayscale difference of pixels in the vertical direction. Then, taking each pixel in the smoothed image as the center, a rectangular set of neighboring pixels that corresponds exactly to the size of the convolution kernel is extracted. Each element of the horizontal convolution kernel is multiplied element-wise with the grayscale value of the corresponding neighboring pixel, and all the multiplication results are summed to obtain the horizontal gradient value of the center pixel. The positive or negative attribute of this value reflects the direction of the horizontal grayscale change, and the absolute value reflects the intensity of the horizontal grayscale change. At the same time, each element of the vertical convolution kernel is multiplied element-wise with the grayscale value of the corresponding neighboring pixel, and all the multiplication results are summed to obtain the vertical gradient value of the center pixel.

[0025] The gradient magnitude of each pixel is calculated using the Euclidean distance formula, which is equal to the square root of the sum of the squares of the horizontal and vertical gradient values. The gradient direction of each pixel is calculated using the arctangent function, which is equal to the arctangent of the ratio of the vertical to the horizontal gradient values. A gradient magnitude histogram is then plotted based on the statistical analysis of the gradient magnitudes of all pixels. The gradient magnitude selection range is determined based on the bimodal distribution of the histogram, and the gradient magnitude corresponding to the second highest peak is taken as the baseline value. This baseline value is then adjusted using the local standard deviation of the smoothed image to obtain an adaptive gradient magnitude threshold. Finally, a number of neighboring pixels matching the average pixel size of the zinc flower are selected, and the angle between the gradient directions of the neighboring pixels and the center pixel is calculated. When the gradient magnitude of the center pixel is greater than the adaptive gradient magnitude threshold, and the gradient directions of the neighboring pixels and the center pixel exhibit a convergence towards or radiating pattern towards the center pixel, this center pixel is considered a candidate center point for the zinc flower.

[0026] Horizontal gradient and vertical gradient refer to the degree of change of gray values ​​in the horizontal and vertical directions, respectively, and are used to describe the brightness change trend of the galvanized surface in different directions. Gradient feature analysis refers to the analysis of brightness change patterns in smooth images based on horizontal and vertical gradients, thereby identifying the locations where brightness changes exhibit radial or converging characteristics. Zinc flower center candidate points refer to the pixel positions detected during gradient feature analysis that are located in the central region of the zinc flower crystal structure. These positions reflect the geometric center characteristics of zinc flower crystal growth and distribution.

[0027] Spatial aggregation of all candidate zinc spangle centers yields several deduplicated zinc spangle centers.

[0028] First, collect the pixel coordinates of all candidate zinc flower centers to construct a complete candidate point coordinate dataset. Iterate through this dataset to calculate the Euclidean distance between every two candidate points. Calculate the arithmetic mean of all Euclidean distances to obtain the average spacing between zinc flower crystals. One-third of this average spacing is determined as the aggregation distance threshold, ensuring that only close candidate points corresponding to the same zinc flower crystal are grouped together. Then, select an unlabeled candidate point from the candidate point coordinate dataset as the initial cluster core. Calculate the Euclidean distance between this initial cluster core and all other unlabeled candidate points in the dataset. Group all unlabeled candidate points whose Euclidean distance is less than the aggregation distance threshold into the same temporary cluster group as the initial cluster core. The process involves grouping and labeling all candidate points within the temporary cluster. Then, a candidate point is selected from the unlabeled candidate points as the new cluster core. This process of distance calculation, candidate point classification, and labeling is repeated until all zinc flower center candidate points in the dataset are included in the corresponding cluster groups and labeled. Finally, for each formed cluster group, the arithmetic mean of the x-coordinate and y-coordinate of the pixel coordinates of all candidate points within the group is calculated. The pixel position corresponding to both the x-coordinate and y-coordinate averages is determined as the final aggregation center of the cluster group. The final aggregation centers of all cluster groups together constitute the set of deduplicated zinc flower center points, effectively eliminating duplicate candidate points caused by the same zinc flower crystal.

[0029] Specifically, spatial aggregation refers to merging candidate points that are close to each other based on their spatial position in the grayscale image, thereby eliminating duplicate detection points caused by the same zinc flower crystal; deduplicated zinc flower center points refer to the set of pixels that can correspond one-to-one with the actual center position of the zinc flower crystal after spatial aggregation, which is used to characterize the spatial distribution characteristics of the zinc flower structure on the surface of the target galvanized steel strip.

[0030] In an embodiment of the present invention, the specific steps for determining the center of the zinc spangle are as follows: Obtain the pixel coordinates of each pixel in a grayscale image; Using the top-left corner of the grayscale image as the origin, the horizontal direction to the right is defined as the positive direction of the x-coordinate, and the vertical direction downward is defined as the positive direction of the y-coordinate. Based on the pixel arrangement pattern of the grayscale image, the total number of columns and rows of the image are counted. For each pixel in the image, its position in the column direction is determined as the x-coordinate value of the pixel coordinate, and its position in the row direction is determined as the y-coordinate value of the pixel coordinate. In this way, a unique two-dimensional coordinate information is assigned to each pixel in the grayscale image.

[0031] Determine the Euclidean distance between the pixel coordinates and the center points of all deduplicated zinc spangles; First, collect the two-dimensional coordinate information of all deduplicated zinc flower center points to construct a set of deduplicated zinc flower center point coordinates. Then, traverse each pixel in the grayscale image and extract the determined two-dimensional pixel coordinates. For each pixel coordinate, calculate the distance between it and each deduplicated zinc flower center point coordinate in the set of deduplicated zinc flower center point coordinates. During the calculation, first calculate the difference between the horizontal coordinate of the pixel and the horizontal coordinate of the deduplicated zinc flower center point, and the difference between the vertical coordinate of the pixel and the vertical coordinate of the deduplicated zinc flower center point. Then, square the two differences respectively, and add the two squared results to get the sum. Finally, calculate the arithmetic square root of the sum. The obtained value is the Euclidean distance between the pixel coordinate and the corresponding deduplicated zinc flower center point. Complete the Euclidean distance calculation between the pixel and all deduplicated zinc flower center points in turn.

[0032] The pixel coordinates of each pixel in a grayscale image refer to the two-dimensional coordinate information used to identify the spatial position of a single pixel in the grayscale image. These coordinates represent the imaging position of the pixel on the surface of the target galvanized steel strip. The Euclidean distance between the pixel coordinates and the center point of the deduplicated zinc flower refers to the distance value in the image coordinate space used to measure the spatial proximity between the pixel position and the center positions of each zinc flower. This distance reflects the relative positional relationship between the pixel on the surface of the galvanized steel strip and the centers of different zinc flower crystals.

[0033] The center point of the deduplicated zinc spangle corresponding to the smallest Euclidean distance is determined as the center of the zinc spangle to which the pixel coordinate belongs.

[0034] The zinc spangle center refers to the zinc spangle center point with the smallest Euclidean distance to a certain pixel coordinate among all deduplicated zinc spangle center points. This zinc spangle center is considered to have a dominant correspondence with the zinc spangle crystal structure of the region where the pixel is located, and is thus used to characterize the zinc spangle crystal region to which the pixel belongs on the surface of the target galvanized steel strip.

[0035] Zinc flower crystals on the surface of the target galvanized steel strip exhibit a spatial structure that grows and expands outward from the crystal center. Different zinc flower crystals occupy relatively independent areas on the surface. Therefore, any pixel must be spatially closest to the center of one of the zinc flower crystals. By calculating the Euclidean distance between the pixel coordinates and the center points of each deduplicated zinc flower, and determining the deduplicated zinc flower center point with the smallest distance as the zinc flower center to which the pixel coordinates belong, the pixel can be assigned to the zinc flower crystal region that is spatially closest to it. This establishes a correspondence between pixels and specific zinc flower crystal structures at the image level.

[0036] Based on the distribution analysis of the candidate stripe regions corresponding to the zinc flower centers, the distribution results are obtained. In an embodiment of the present invention, the specific steps for obtaining the distribution result are as follows: One-dimensional difference processing is performed on the pixel gray levels of a grayscale image to obtain a gray level change sequence; In detail, firstly, according to the spatial arrangement order of the grayscale image, adjacent pixels in the same row or column of the grayscale image are selected as processing objects to ensure that the pixels have continuity in spatial position; then, the grayscale values ​​of adjacent pixels are obtained in sequence, and the grayscale values ​​of two adjacent pixels are subtracted to obtain the grayscale change at the corresponding position; all grayscale changes obtained along the same direction are arranged in pixel order to form a grayscale change sequence that describes the grayscale change with spatial position.

[0037] Pixel grayscale refers to the numerical value in a grayscale image used to characterize the brightness intensity of a corresponding position on the surface of a target galvanized steel strip. This value reflects the reflection and scattering state of the galvanized surface to light. One-dimensional difference processing refers to performing difference operations on the grayscale values ​​of adjacent pixels along a predetermined direction to obtain the grayscale variation with spatial position. Grayscale change sequence refers to the sequence data formed after one-dimensional difference processing, used to describe the amplitude and trend of grayscale change, reflecting the texture undulation characteristics of the galvanized steel strip surface in the corresponding direction.

[0038] The grayscale image is divided into regions based on the grayscale change sequence to obtain several stripe candidate regions; The grayscale change sequence is mapped one-to-one with the pixel positions in the grayscale image, so that each grayscale change can be mapped back to a specific spatial position in the grayscale image. Then, the continuity of the grayscale changes is analyzed along the arrangement direction of the grayscale change sequence. When the grayscale changes of adjacent positions show a continuous change relationship in space, the corresponding pixel positions are divided into the same region. When the grayscale changes show obvious separation in space, the region is segmented at the corresponding position. By traversing and analyzing the entire grayscale change sequence, the division of multiple continuous regions in the grayscale image is completed, thereby obtaining several candidate stripe regions that show a banded distribution in spatial morphology.

[0039] Stripe candidate regions refer to image regions in a grayscale image that are divided according to the grayscale change sequence and have continuous or periodic grayscale changes. These regions exhibit a band-like structure in spatial morphology and are areas that may have stripe features.

[0040] Based on the zinc flower centers corresponding to all pixel coordinates within the stripe candidate region, determine the subset of zinc flower centers of the stripe candidate region. First, each stripe candidate region is traversed pixel by pixel in row and column order, and the pixel coordinates of each pixel in the region are extracted one by one. Then, according to the pre-determined correspondence between pixels and their respective zinc flower centers, the coordinates of the zinc flower center corresponding to each pixel coordinate are obtained. Subsequently, an empty coordinate set is initialized, and all the extracted zinc flower center coordinates are added to the empty set one by one. The characteristics of the set are used to automatically filter duplicate coordinate data. The final coordinate set is the zinc flower center subset of the stripe candidate region.

[0041] Determine the azimuth angle between the pixel coordinates and the center of its corresponding zinc flower; For each pixel within the stripe candidate region, first extract the pixel coordinates and the coordinates of its corresponding zinc flower center. Calculate the difference between the pixel's horizontal coordinate and the horizontal coordinate of its corresponding zinc flower center as the horizontal component, and calculate the difference between the pixel's vertical coordinate and the vertical coordinate of its corresponding zinc flower center as the vertical component. Use the arctangent function to calculate the ratio of the vertical component to the horizontal component to obtain the initial radian value. Multiply the initial radian value by 180 and divide by pi (π) to complete the conversion from radians to degrees, obtaining the initial angle value. Finally, determine the initial angle value based on the positive and negative attributes of the horizontal and vertical components. The azimuth angle is determined by the quadrant. If both the horizontal and vertical components are positive, the pixel is in the first quadrant, and the angle value remains unchanged. If both the horizontal and vertical components are negative, the pixel is in the second quadrant, and the initial angle value is increased by 180 degrees. If both the horizontal and vertical components are negative, the pixel is in the third quadrant, and the initial angle value is increased by 180 degrees. If both the horizontal and vertical components are positive, the pixel is in the fourth quadrant, and the initial angle value is increased by 360 degrees. This results in an azimuth angle within the range of 0 to 360 degrees. The azimuth angle accurately depicts the relative directional position of the pixel within the corresponding zinc flower crystal structure.

[0042] The zinc flower center subset refers to the set of all zinc flower centers corresponding to all pixels within the stripe candidate region, used to describe the spatial relationship between the stripe candidate region and multiple zinc flower crystals; the azimuth angle refers to the direction angle of the pixel coordinates relative to its corresponding zinc flower center in the image plane, used to characterize the relative directional position of the pixel in the zinc flower crystal structure.

[0043] The azimuth distribution of the stripe candidate region is determined based on the azimuth angles corresponding to the coordinates of all pixels within the stripe candidate region. The stripe candidate region is traversed pixel by pixel, and the azimuth data corresponding to all pixels in the region are collected to form a complete azimuth dataset. The total number of azimuths in the dataset and the maximum and minimum values ​​are counted. The difference between the maximum and minimum values ​​is calculated to obtain the azimuth distribution span. The number of angle intervals is determined according to the total number of azimuths. The more azimuths there are, the more intervals are divided to ensure statistical accuracy. The complete angle range from 0 degrees to 360 degrees is equally divided according to the determined number of intervals to obtain several continuous and non-overlapping angle intervals. Then, the number of azimuth data contained in each angle interval is counted, and a mapping relationship between each angle interval and the corresponding number of azimuths is established. This mapping relationship is the azimuth distribution of the stripe candidate region, which fully reflects the consistency or discreteness characteristics of pixels in the region in the spatial direction.

[0044] The distribution of the stripe candidate region was analyzed by azimuth angle distribution, and the distribution results were obtained.

[0045] First, all azimuth data within the stripe candidate region are converted into corresponding unit vectors. The x-coordinate of each azimuth unit vector is calculated using the cosine function of that azimuth, and the y-coordinate is calculated using the sine function of that azimuth. Then, the x-coordinates of all unit vectors are summed to obtain the total x-coordinate sum, and the y-coordinates of all unit vectors are summed to obtain the total y-coordinate sum. The square root of the sum of the x-coordinate sum and the square root of the sum of the y-coordinate sum is taken to obtain the length of the composite vector. This value directly reflects the overall concentration of the azimuth. Simultaneously, the circular square deviation is calculated, which is equal to 1 minus the composite vector length. A larger circular square deviation indicates a more dispersed azimuth distribution. Next, the ratio of the sum of the x-coordinates to the sum of the y-coordinates is calculated using the arctangent function to obtain the principal direction angle, i.e., the direction of concentration of the azimuth distribution. Finally, the circular standard deviation is calculated, which is equal to the square root of -2 multiplied by the natural logarithm. The product of the composite vector length and the azimuth angle length is used to quantify the discrete range of the azimuth angle relative to the principal direction. Then, a concentrated angle interval is defined with the principal direction angle as the center and twice the standard deviation of the circle as the range. The number of azimuth angles contained in this concentrated angle interval is counted, and its proportion of the total number of azimuth angles is calculated. At the same time, the circular variance of all azimuth angles is statistically analyzed to obtain the arithmetic mean and standard deviation of the circular variance. The difference between the arithmetic mean and standard deviation of the circular variance is used as the concentration judgment threshold. If the composite vector length is greater than 1 minus the concentration judgment threshold, and the proportion of the number of azimuth angles in the concentrated angle interval is greater than the composite vector length, then it is determined that the azimuth angles are concentrated in a single principal direction. If the composite vector length is less than or equal to 1 minus the concentration judgment threshold, or the proportion of the number of azimuth angles in the concentrated angle interval is less than or equal to the composite vector length, then it is determined that the azimuth angles are not concentrated in a single principal direction.

[0046] Azimuth distribution refers to the angular distribution state obtained by summing the azimuth angles formed by all pixel coordinates relative to the center of their respective zinc flower within the stripe candidate region. This distribution is used to reflect the consistency or dispersion characteristics of pixels in the spatial direction within the stripe candidate region. The distribution result refers to the analytical output obtained by statistically analyzing the angular distribution state based on the azimuth distribution. This output is used to characterize whether the stripe candidate region exhibits concentrated, dispersed, or multi-directional distribution characteristics in the spatial direction, thereby providing a basis for subsequent classification of the stripe candidate region.

[0047] Since the zinc flower texture on the surface of the target galvanized steel strip appears as a spatial structure extending outward from the center of the zinc flower crystal, and the pseudo-defect stripes formed by sampling aliasing are an image phenomenon generated by the interaction between the zinc flower spatial structure and the fixed sampling grid, these stripes often show similar directions and consistent orientations in multiple zinc flower regions. By statistically analyzing the azimuth angles of pixels in the candidate stripe region relative to their respective zinc flower centers to form an azimuth angle distribution, and analyzing this distribution, it can be determined whether the stripes have the characteristic of being concentrated in a single main direction in the spatial direction. This distinguishes whether the stripes are pseudo-defect textures caused by the superposition of the zinc flower crystal spatial structure and the sampling phase, and avoids misjudging such artifacts as real material defects based solely on local grayscale changes.

[0048] The pseudo-defect stripes formed by sampling aliasing refer to the phenomenon that, when imaging a high-speed galvanized steel strip online, the zinc flower crystal texture on the galvanized surface has a high spatial frequency distribution. The imaging process uses a fixed sampling interval and a fixed scanning trajectory for discrete sampling. Under specific sampling phase conditions, the high-frequency zinc flower texture is folded and appears as a low-frequency strip-shaped brightness change. This stripe presents a morphological feature similar to a real stripe defect in the grayscale image, but there is no corresponding material defect on the actual galvanized steel strip. Its essence is a non-real defect texture phenomenon generated at the image level by the interaction between the spatial structure of the zinc flower crystal and the sampling grid.

[0049] Virtual resampling analysis is performed on the stripe candidate region to determine the sampling phase sensitivity measure; In an embodiment of the present invention, the specific steps for determining the sampling phase sensitivity metric are as follows: Determine the target location of the stripe candidate region in the grayscale image; First, traverse the stripe candidate region pixel by pixel, recording the x-coordinate and y-coordinate of each pixel in the grayscale image. Count all recorded x-coordinate values ​​to obtain the minimum and maximum x-coordinate values. At the same time, count all recorded y-coordinate values ​​to obtain the minimum and maximum y-coordinate values. Then, use the minimum x-coordinate and minimum y-coordinate values ​​as the coordinates of the top-left vertex, and use the maximum x-coordinate and maximum y-coordinate values ​​as the coordinates of the bottom-right vertex to determine the smallest rectangular region containing all pixels of the stripe candidate region. The coordinate range of the rectangular region is the target position of the stripe candidate region in the grayscale image.

[0050] Local resampling is performed on the target location to obtain several virtual resampled image patches; First, the region to be resampled is determined as the rectangular region corresponding to the target location. The number of pixel columns and rows in this region is counted to determine the spatial density of the original sampling. Based on the pixel interval of the original sampling, the offset range of resampling is determined. The offset value is a different value between 0 and the original pixel interval, and the values ​​are not repeated. For each offset, the offset is superimposed on the horizontal axis to obtain the new sampling starting point horizontal coordinate, and superimposed on the vertical axis to obtain the new sampling starting point vertical coordinate. Then, based on the new sampling starting point, the pixel gray values ​​of the corresponding positions are extracted row by row and column by column according to the number of pixel columns and rows of the original region to form an image sub-block with the same size as the target location region. All image sub-blocks obtained by different offsets together constitute a set of virtual resampled image blocks.

[0051] The target position of the stripe candidate region in a grayscale image refers to the spatial coordinate range of the stripe candidate region in the grayscale image, used to identify the specific position of the stripe candidate region in the imaging of the target galvanized steel strip surface; local resampling refers to re-acquiring pixel values ​​of the image region corresponding to the target position according to different sampling starting points or sampling alignment methods while keeping the spatial range of the stripe candidate region unchanged, thereby simulating the imaging result of the region when the imaging sampling conditions change; virtual resampled image blocks refer to multiple image sub-blocks obtained through local resampling. These image sub-blocks correspond to the same stripe candidate region in spatial position, but differ in sampling starting points or sampling alignment relationships, used to reflect the impact of changes in sampling conditions on the stripe imaging effect.

[0052] A one-dimensional projection of the pixel grayscale values ​​of the virtual resampled image patch yields a stripe intensity sequence. First, a projection direction consistent with the stripe extension direction is determined in each virtual resampled image block, so that the projection direction is consistent with the main direction of the stripes in the image. Then, along the projection direction, the gray values ​​of pixels in the same straight line or column in the virtual resampled image block are accumulated or averaged, so that the pixel gray information distributed in two-dimensional space is compressed and mapped into a one-dimensional gray sequence arranged along the projection direction. By traversing the entire virtual resampled image block, a complete one-dimensional gray change sequence is obtained. This sequence can reflect the brightness fluctuation of the stripes in the corresponding direction, thus forming a stripe intensity sequence.

[0053] The variance of the stripe intensity sequence is used as the stripe significance value; The average level of all gray values ​​in the stripe intensity sequence is calculated by statistically summing each gray value. Then, the deviation of each gray value from the average level is calculated, and each deviation is squared. All squared deviations are summed and normalized to obtain a numerical result reflecting the overall dispersion of the stripe intensity sequence. This numerical result characterizes the salience of the stripes in the virtual resampled image patch. The greater the dispersion of the stripe intensity sequence, the greater the corresponding stripe salience value, thus providing a clear data basis for subsequent sampling phase sensitivity analysis of stripe candidate regions.

[0054] One-dimensional projection refers to the summarization and processing of pixel grayscale values ​​in a virtual resampled image block along a predetermined direction, transforming two-dimensional image information into a one-dimensional grayscale sequence describing the intensity variation of stripes; the stripe intensity sequence refers to the grayscale variation sequence obtained through one-dimensional projection, which is used to characterize the brightness fluctuation of stripes in the corresponding direction; the stripe salience value refers to the numerical result obtained after statistical calculation of the stripe intensity sequence, which reflects the degree of identifiability of stripes in the image.

[0055] The sampling phase sensitivity measure of the stripe candidate region is determined based on the stripe saliency values ​​corresponding to all virtual resampled image patches.

[0056] For the same fringe candidate region, all fringe saliency values ​​obtained under different virtual resampling conditions are collected, so that each fringe saliency value corresponds to the imaging result of the fringe candidate region under a certain sampling phase condition. Then, the fringe saliency values ​​are statistically analyzed as a whole. By comparing the degree of difference between each fringe saliency value, the fluctuation of fringe saliency with the change of sampling phase is characterized. Furthermore, the dispersion of the fringe saliency values ​​is quantitatively calculated so that the dispersion can reflect the magnitude of the change of fringe saliency under different sampling phase conditions. Finally, the dispersion is used as a measure of the sampling phase sensitivity of the fringe candidate region.

[0057] The sampling phase sensitivity measure is a comprehensive measure based on the changes in the saliency values ​​of stripes corresponding to multiple virtual resampled image patches. It is used to characterize the response of stripe candidate regions to changes in the sampling start point, thereby reflecting whether the stripe is prone to significant imaging differences due to changes in sampling conditions.

[0058] The defect labels corresponding to the grayscale images are updated based on the sampling phase sensitivity measurement and distribution results to obtain updated labels; In an embodiment of the present invention, the specific steps for obtaining the updated tag are as follows: The candidate regions of the stripes are grouped according to the sampling phase sensitivity metric to obtain the grouping results; First, for all stripe candidate regions in the same grayscale image, the sampling phase sensitivity measure corresponding to each stripe candidate region is obtained to make the stripe candidate regions comparable in terms of sampling phase sensitivity. Then, the sampling phase sensitivity measures of all stripe candidate regions are sorted as a whole, forming an ordered arrangement from largest to smallest. In the ordered arrangement, stripe candidate regions with sampling phase sensitivity measures at the top of the sort are divided into a first-class sensitive group, which represents stripe candidate regions with a high degree of response to sampling phase changes, and stripe candidate regions with sampling phase sensitivity measures at the bottom of the sort are divided into a second-class sensitive group, which represents stripe candidate regions with a low degree of response to sampling phase changes. Through the above division method based on the relative size relationship of sampling phase sensitivity measures, the stripe candidate regions are grouped to obtain a grouping result including the first-class sensitive group and the second-class sensitive group.

[0059] After obtaining the sampling phase sensitivity metrics corresponding to all stripe candidate regions and arranging them in descending order, the difference between two adjacent sampling phase sensitivity metrics is calculated one by one to obtain a difference sequence reflecting the sensitivity change amplitude of adjacent stripe candidate regions. In the difference sequence, the position where the difference reaches its maximum is determined, and the two adjacent sampling phase sensitivity metrics corresponding to the maximum difference are taken as the boundary position where the sensitivity change is most significant. The median value between the sampling phase sensitivity metrics located on both sides of the boundary position is used as the boundary value to distinguish the degree of sampling phase sensitivity. Stripe candidate regions with sampling phase sensitivity metrics greater than the boundary value are determined to be stripe candidate regions with a high degree of response to sampling phase changes. Stripe candidate regions with sampling phase sensitivity metrics less than or equal to the boundary value are determined to be stripe candidate regions with a low degree of response to sampling phase changes.

[0060] Based on the grouping and distribution results, the candidate stripe regions are classified to obtain the classification results. Specifically, the grouping results refer to the classification information after dividing the candidate stripe regions into several categories according to the sampling phase sensitivity metric corresponding to each candidate stripe region. This grouping is used to distinguish candidate stripe regions with different degrees of response to sampling phase changes. The classification results refer to the stripe type judgment result determined for each candidate stripe region based on the combination of the grouping and distribution results. This is used to indicate whether the candidate stripe region is closer to the pseudo-defect stripe formed by sampling aliasing or closer to the real material defect stripe.

[0061] In an embodiment of the present invention, the specific steps for obtaining the category determination result are as follows: If the grouping result of the stripe candidate region is a sensitive group and the distribution result is that the azimuth distribution satisfies the distribution characteristic of the azimuth angle concentrated in a single main direction, then the category determination result of the stripe candidate region is set as the zinc flower sampling aliasing pseudo-defect stripe region. The candidate region of this stripe is identified as a pseudo-defect stripe region due to the dual characteristics of being dominated by both the sampling mechanism and the spatial structure of the zinc flower. On the one hand, a sensitive group indicates that the saliency of the stripe candidate region fluctuates significantly with changes in the sampling phase, suggesting that the imaging performance of the stripe is highly dependent on the sampling starting point, which is consistent with the basic characteristics of sampling aliasing. On the other hand, the azimuth distribution is concentrated in a single principal direction, indicating that the stripe exhibits a consistent orientation pattern in multiple zinc flower regions. This orientation is not randomly generated by local material anomalies, but rather has a stable geometric coupling relationship with the spatial structure of the zinc flower crystal expanding outward from the center. When both of these characteristics occur simultaneously, it can be reasonably determined that the stripe does not originate from a real material defect on the surface of the target galvanized steel strip, but is an image-level pseudo-defect texture formed by the interaction between the spatial structure of the zinc flower crystal and the fixed sampling grid under specific sampling phase conditions.

[0062] If the grouping result of the stripe candidate region is a Class II sensitive group, and the distribution result is an azimuth distribution whose distribution characteristics do not satisfy the requirement that the azimuth angle is concentrated in a single main direction, then the category determination result of the stripe candidate region is set as a real stripe defect region.

[0063] The candidate region of this stripe was identified as a real stripe defect region because this type of stripe showed a low response to changes in the sampling phase during imaging, indicating that its stripe salience remained relatively stable under different sampling starting conditions and did not fluctuate significantly depending on changes in the sampling grid position. This characteristic is consistent with the stable performance of real material defects during imaging. The azimuth distribution is not concentrated in a single principal direction, indicating that the stripes lack a geometric coupling relationship consistent with the zinc flower crystal structure in the spatial direction. Their orientation shows a more dispersed or irregular distribution state. This distribution characteristic is more consistent with the spatial characteristics of real stripe defects that are randomly generated by material surface anomalies or process factors.

[0064] If the aforementioned two judgment conditions are not met, that is, if the stripe candidate region is neither a type I sensitive group and the azimuth angle is concentrated in a single main direction, nor a type II sensitive group and the azimuth angle is not concentrated in a single main direction, it indicates that the stripe candidate region exhibits an uncertain or contradictory state in terms of sampling phase response characteristics or spatial direction distribution characteristics. In this state, it is impossible to make a clear judgment on its cause based solely on the existing sampling phase sensitivity measurement and azimuth angle distribution results. Therefore, the category judgment result of the stripe candidate region is set as a stripe region to be judged, indicating that the region can neither be reliably classified as a pseudo-defect stripe formed by zinc spangle sampling aliasing, nor can it be reliably judged as a real stripe defect.

[0065] Obtain the defect labels corresponding to the grayscale images; First, a unique identifier for the grayscale image is determined. This identifier consists of the total number of rows and columns of pixels in the grayscale image, the imaging time, and the production batch number and roll number of the corresponding target galvanized steel strip. Then, the original inspection process record bound to this unique identifier is searched, and information associated with the grayscale image is extracted from the record, including the defect type of the identifier, the coordinates of the local area of ​​the grayscale image corresponding to the defect type, and the execution time of the inspection operation. This extracted information is organized according to the way that the defect type is associated with the corresponding area coordinates. The organized information set is the defect label corresponding to the grayscale image.

[0066] The defect labels corresponding to the grayscale images are updated based on the category determination results to obtain updated labels.

[0067] The classification results corresponding to the stripe candidate regions are obtained, and the classification results are matched one-to-one with the specific spatial regions in the grayscale image. Then, the classification results are compared with the original defect labels of the stripe candidate regions. When the classification results are inconsistent with the original defect labels, the defect category identifier of the corresponding stripe candidate region in the original defect label is replaced with the classification results. When the classification results are consistent with the original defect labels, the original defect labels are kept unchanged. By performing the above update operation on all stripe candidate regions in the grayscale image one by one, the defect labels corresponding to the grayscale image as a whole reflect the classification results after sampling phase sensitivity analysis and azimuth distribution analysis, thereby obtaining the updated defect labels, which are used to generate accurate and reliable defect corpora in the future.

[0068] Specifically, the defect label corresponding to the grayscale image refers to the defect category identification information associated with the grayscale image of the target galvanized steel strip. It is used to record the defect type corresponding to the grayscale image or its local area in the original detection process. The updated label refers to the new defect category identification information formed after adjusting and correcting the original defect label after obtaining the category determination result of the stripe candidate area. The updated label reflects the true stripe attributes obtained after sampling phase sensitivity analysis and distribution analysis. It is used to ensure that the defect data generated subsequently can accurately correspond to the actual state of the target galvanized steel strip surface, thereby avoiding the impact of sampling aliased false defect stripes on the quality of the defect data.

[0069] Because in the original grayscale image corpus, the defect labels of stripe candidate regions are usually generated based on conventional appearance features or traditional detection rules, they cannot distinguish between real material defects and sampled aliasing pseudo-defect stripes generated by the superposition of zinc flower crystal spatial structure and sampling phase. After introducing analysis of the spatial distribution characteristics of zinc flower and the sensitivity of sampling phase, it is possible to clearly identify which stripe candidate regions belong to image-level pseudo-defect textures and which belong to real stripe defects on the surface of the target galvanized steel strip. If the original defect labels are still used, sampled aliasing pseudo-defect stripes will continue to be mixed into the real defect corpus, thus affecting subsequent corpus-based model training, defect statistical analysis, and knowledge construction. Updating the defect labels corresponding to the grayscale images according to the category determination results, so that the label content truly reflects the cause attributes of the stripes, is a necessary step to ensure that the generated defect corpus is consistent with the actual surface state of the target galvanized steel strip and to avoid pseudo-defects contaminating the AI ​​corpus.

[0070] A corpus of defective AI data generated from updated labels and grayscale images of target galvanized steel strips.

[0071] In an embodiment of the present invention, the specific steps for generating a defective AI corpus set are as follows: By updating the labels, all stripe candidate regions are classified to obtain the first stripe region set, the second stripe region set, and the third stripe region set; According to the category meaning indicated by the updated labels, the stripe candidate regions are traversed and classified one by one. When the updated label of a stripe candidate region indicates that it is a pseudo-defect stripe formed by zinc spangle sampling aliasing, the stripe candidate region is assigned to the first stripe region set. When the updated label of a stripe candidate region indicates that it is a real stripe defect, the stripe candidate region is assigned to the second stripe region set. When the updated label of a stripe candidate region indicates that it is in a pending state, the stripe candidate region is assigned to the third stripe region set. By performing the above classification operation on all stripe candidate regions, each stripe candidate region is uniquely assigned to the corresponding region set only according to its updated label, thereby completing the classification of stripe candidate regions and obtaining the first stripe region set, the second stripe region set, and the third stripe region set that are distinguishable and non-overlapping, providing a clear data foundation for the subsequent construction of a real defect corpus.

[0072] The first stripe region set refers to the set of candidate stripe regions that are determined to be zinc spangle sampling aliasing false defect stripes based on the updated label; the second stripe region set refers to the set of candidate stripe regions that are determined to be real stripe defects based on the updated label; and the third stripe region set refers to the set of candidate stripe regions that are determined to be stripes to be judged based on the updated label.

[0073] Obtain the coil index of the target galvanized steel strip; First, the unique identifier information of the grayscale image corresponding to the target galvanized steel strip is extracted. This identifier information is composed of the physical number of the imaging device, the precise timestamp of the imaging operation, and the frame number of the grayscale image in the imaging sequence. Then, using this unique identifier information as the query basis, a data association query request is initiated to the cloud platform storing the entire production process data of galvanized steel strip. After receiving the request, the cloud platform first matches the corresponding production line number based on the physical number of the imaging device in the identifier information. Then, using the precise timestamp in the identifier information, it locates the production batch number, coil number, production date, and production line station number of the target galvanized steel strip in the real-time production record corresponding to the production line number. Subsequently, according to the combination rules specified in the galvanized steel strip production data management specification, the production batch number, coil number, production date, and production line station number are sequentially concatenated to form a string that can uniquely identify the coil identity of the target galvanized steel strip. Finally, the cloud platform feeds back this string to the requesting end, and this string is the coil index of the target galvanized steel strip.

[0074] Based on the second stripe region set, grayscale images, and roll material index, a set of real defect corpora is constructed; The coil index refers to the information used to uniquely identify the target galvanized steel strip coil, which is used to associate grayscale images and their corresponding stripe areas with specific production coils; the real defect corpus refers to the data set constructed based on the second stripe area set, grayscale images, and the coil index. This set contains image areas that can truly reflect the existence of material defects on the surface of the target galvanized steel strip and their corresponding identification information.

[0075] The set of real defect corpora is used as the set of defect AI corpora for the target galvanized steel strip.

[0076] First, retrieve all stripe candidate region information from the second stripe region set. Extract the minimum and maximum values ​​of the horizontal and vertical coordinates for each stripe candidate region. Based on these coordinate values, extract a rectangular pixel block containing the stripe candidate region from the corresponding grayscale image as a defect image sub-block. Simultaneously, extract defect feature information such as the mean grayscale value, grayscale variance, azimuth distribution, principal direction angle, circular variance, and stripe saliency value within this sub-block. Then, using the coil index as the query key, initiate a request to the galvanized steel strip production data module of the cloud platform to obtain the corresponding production process parameters for the coil, including coating thickness, rolling speed, annealing temperature, and production station number. The coordinate information of the candidate stripe regions of the defective image sub-blocks, defect feature information, roll material index, and production process parameters obtained from the cloud platform are correlated and integrated to form a single real defect corpus data. After the corpus integration of all candidate stripe regions in the second stripe region set is completed in sequence, all single corpus data are uploaded to the defect corpus storage module of the cloud platform. The cloud platform automatically verifies the integrity of the image format data of each corpus and performs deduplication processing through the combination of roll material index and region coordinate identifier to ensure that each corpus is unique and the information is complete. Finally, all corpus data that pass the verification are summarized into a dataset, which is the real defect corpus set.

[0077] A defect AI corpus refers to a data set consisting of real defect corpora used for subsequent defect identification, analysis, or training. Its content only includes corpus data consistent with the real defect state of the target galvanized steel strip.

[0078] An embodiment of the present invention also provides an AI corpus processing system for the metal products industry based on a cloud platform.

[0079] In this embodiment, the functions of each module / unit are as follows: The zinc spangle center determination module is used to extract the zinc spangle center from the grayscale image of the target galvanized steel strip and obtain the deduplicated zinc spangle center point; based on the deduplicated zinc spangle center point, the zinc spangle center to which the pixel coordinates in the grayscale image belong is determined. The stripe feature analysis module is used to perform distribution analysis on the stripe candidate regions corresponding to the grayscale image based on the center of the zinc flower, and obtain the distribution results; it also performs virtual resampling analysis on the stripe candidate regions to determine the sampling phase sensitivity measure. The defect label update module is used to update the defect labels corresponding to the grayscale image based on the sampling phase sensitivity measurement and distribution results, and obtain the updated labels. The AI ​​corpus generation module is used to generate a defect AI corpus set of target galvanized steel strips based on updated labels and grayscale images.

[0080] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A cloud-based AI corpus processing method for the metal products industry, characterized in that, Includes the following steps: The center of the zinc flower is extracted from the grayscale image of the target galvanized steel strip to obtain the center point of the deduplicated zinc flower; the center of the zinc flower to which the pixel coordinates in the grayscale image belong is determined based on the center point of the deduplicated zinc flower. Based on the distribution analysis of the candidate stripe regions corresponding to the zinc flower centers, the distribution results are obtained. Virtual resampling analysis is performed on the stripe candidate region to determine the sampling phase sensitivity measure; The defect labels corresponding to the grayscale images are updated based on the sampling phase sensitivity measurement and distribution results to obtain updated labels; A corpus of defective AI data generated from updated labels and grayscale images of target galvanized steel strips.

2. The AI ​​corpus processing method for the metal products industry based on a cloud platform according to claim 1, characterized in that, The specific steps to obtain the center point of the de-zinc spangle are as follows: Obtain the grayscale image of the target galvanized steel strip; A smoothed image is obtained by smoothing a grayscale image; Based on the horizontal and vertical gradients of the smoothed image, gradient feature analysis is performed on the smoothed image to obtain several candidate points for the center of zinc flower. Spatial aggregation of all candidate zinc spangle centers yields several deduplicated zinc spangle centers.

3. The AI ​​corpus processing method for the metal products industry based on a cloud platform according to claim 1, characterized in that, The specific steps to determine the zinc flower center are as follows: Obtain the pixel coordinates of each pixel in a grayscale image; Determine the Euclidean distance between the pixel coordinates and the center points of all deduplicated zinc spangles; The center point of the deduplicated zinc spangle corresponding to the smallest Euclidean distance is determined as the center of the zinc spangle to which the pixel coordinate belongs.

4. The AI ​​corpus processing method for the metal products industry based on a cloud platform according to claim 3, characterized in that, The specific steps to obtain the distribution results are as follows: One-dimensional difference processing is performed on the pixel gray levels of a grayscale image to obtain a gray level change sequence; The grayscale image is divided into regions based on the grayscale change sequence to obtain several stripe candidate regions; Based on the zinc flower centers corresponding to all pixel coordinates within the stripe candidate region, determine the subset of zinc flower centers of the stripe candidate region. Determine the azimuth angle between the pixel coordinates and the center of its corresponding zinc flower; The azimuth distribution of the stripe candidate region is determined based on the azimuth angles corresponding to the coordinates of all pixels within the stripe candidate region. The distribution of the stripe candidate region was analyzed by azimuth angle distribution, and the distribution results were obtained.

5. The AI ​​corpus processing method for the metal products industry based on a cloud platform according to claim 1, characterized in that, The specific steps for determining the sampling phase sensitivity metric are as follows: Determine the target location of the stripe candidate region in the grayscale image; Local resampling is performed on the target location to obtain several virtual resampled image patches; A one-dimensional projection of the pixel grayscale values ​​of the virtual resampled image patch yields a stripe intensity sequence. The variance of the stripe intensity sequence is used as the stripe significance value; The sampling phase sensitivity measure of the stripe candidate region is determined based on the stripe saliency values ​​corresponding to all virtual resampled image patches.

6. The AI ​​corpus processing method for the metal products industry based on a cloud platform according to claim 1, characterized in that, The specific steps to obtain the updated tags are as follows: The candidate regions of the stripes are grouped according to the sampling phase sensitivity metric to obtain the grouping results; Based on the grouping and distribution results, the candidate regions of the stripes are classified, and the classification results are obtained. Obtain the defect labels corresponding to the grayscale images; The defect labels corresponding to the grayscale images are updated based on the category determination results to obtain updated labels.

7. The AI ​​corpus processing method for the metal products industry based on a cloud platform according to claim 6, characterized in that, The specific steps to obtain the category determination result are as follows: If the grouping result of the stripe candidate region is a sensitive group and the distribution result is that the azimuth distribution satisfies the distribution characteristic of the azimuth angle concentrated in a single main direction, then the category determination result of the stripe candidate region is set as the zinc flower sampling aliasing pseudo-defect stripe region. If the grouping result of the stripe candidate region is a Class II sensitive group, and the distribution result is an azimuth distribution whose distribution characteristics do not satisfy the requirement that the azimuth angle is concentrated in a single main direction, then the category determination result of the stripe candidate region is set as a real stripe defect region.

8. The AI ​​corpus processing method for the metal products industry based on a cloud platform according to claim 1, characterized in that, The specific steps for generating the defective AI corpus are as follows: By updating the labels, all stripe candidate regions are classified to obtain the first stripe region set, the second stripe region set, and the third stripe region set; Obtain the coil index of the target galvanized steel strip; Based on the second stripe region set, grayscale images, and roll material index, a set of real defect corpora is constructed; The set of real defect corpora is used as the set of defect AI corpora for the target galvanized steel strip.

9. A cloud-based AI corpus processing system for the metal products industry, characterized in that: The system includes: The zinc spangle center determination module is used to extract the zinc spangle center from the grayscale image of the target galvanized steel strip and obtain the deduplicated zinc spangle center point; based on the deduplicated zinc spangle center point, the zinc spangle center to which the pixel coordinates in the grayscale image belong is determined. The stripe feature analysis module is used to perform distribution analysis on the stripe candidate regions corresponding to the grayscale image based on the center of the zinc flower, and obtain the distribution results; it also performs virtual resampling analysis on the stripe candidate regions to determine the sampling phase sensitivity measure. The defect label update module is used to update the defect labels corresponding to the grayscale image based on the sampling phase sensitivity measurement and distribution results, and obtain the updated labels. The AI ​​corpus generation module is used to generate a defect AI corpus set of target galvanized steel strips based on updated labels and grayscale images.

Images