Continuous casting billet surface image defect detection method based on image processing

Abstract

The invention relates to the technical field of continuous casting billet detection, and discloses a continuous casting billet surface image defect detection method based on image processing, and the method comprises the following steps: collecting a continuous casting billet surface multispectral image sequence, separating an oxide scale and a matrix metal area by means of spectral reflectivity difference, and building a surface material distribution map; constructing a dynamic threshold segmentation model, generating an adaptive segmentation interval according to gray value fluctuation characteristics, and extracting a contour boundary of a potential defect area; establishing a defect morphological feature library, and screening candidate defect areas through multi-scale matching and geometric constraint; performing spectrum verification, and comparing the similarity between the reflectivity mutation curve and a preset defect characteristic curve; generating a three-dimensional shape reconstruction instruction, and triggering a laser displacement sensor to acquire elevation data; and establishing a defect classification decision tree, inputting elevation and multispectral features, and outputting a final defect category according to a material type and deformation degree combination rule.

Description

Technical Field

[0001] This invention relates to the field of continuous casting billet inspection technology, specifically to a method for detecting surface image defects in continuous casting billets based on image processing. Background Technology

[0002] In the steel production process, continuously cast billets are an important intermediate product, and their surface quality directly affects the performance and quality of subsequent rolled products. Currently, the detection of surface defects in continuously cast billets mainly relies on manual visual inspection and traditional physical testing methods. During manual visual inspection, inspectors need to work continuously in a high-temperature, high-dust production environment. Prolonged observation can easily lead to visual fatigue, making it difficult to consistently identify minute defects. Furthermore, the judgment standards of different inspectors vary, which can easily lead to missed or false detections, making it impossible to form unified and standardized test results.

[0003] In traditional physical testing methods, some methods employ contact testing, where the testing device must directly contact the surface of the continuously cast billet. Under the high temperature of the billet, this not only easily causes wear and tear on the testing device, shortening its service life, but may also cause secondary damage to the surface of the billet during the contact process. Other non-contact testing methods use single-spectrum image acquisition and analysis, which can only obtain limited optical information from the surface of the billet. It is difficult to effectively distinguish between the area covered by iron oxide scale and the base metal area. As a result, in the defect identification stage, the natural texture of iron oxide scale is easily misjudged as a defect, or the true defect is missed because the subtle anomalies in the base metal area are not accurately captured.

[0004] As the steel industry continues to demand higher production efficiency and product quality, the limitations of traditional testing methods are becoming increasingly apparent. On the one hand, traditional methods have low testing efficiency, making it difficult to match the continuous rhythm of continuous casting production. Testing delays are common, and if defects are not detected in time, substandard continuously cast billets will enter subsequent processes, increasing processing costs and wasting resources. On the other hand, traditional methods lack the ability to classify and analyze the causes of defects. They can only make a preliminary judgment on the existence of defects, failing to accurately distinguish defect types and infer their causes. This hinders targeted adjustments to process parameters during production, making it difficult to reduce defects at the source and restricting the overall improvement of steel production quality. Summary of the Invention

[0005] The purpose of this invention is to provide a method for detecting surface image defects of continuously cast billets based on image processing, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for detecting surface image defects of continuously cast billets based on image processing, the method comprising:

[0007] A multispectral image sequence of the surface of the continuously cast billet was acquired, and the area covered by iron oxide scale and the base metal area were separated by the difference in spectral reflectance to establish a surface material distribution map.

[0008] A dynamic threshold segmentation model is constructed to generate adaptive segmentation intervals based on the gray value fluctuation characteristics of different regions in the material distribution map, and to extract the contour boundaries of potential defect regions.

[0009] Establish a defect morphology feature library, perform multi-scale matching between the contour boundary and the standard defect template, and filter out candidate defect regions that meet the geometric constraints.

[0010] Perform spectral verification of the defect region and compare the similarity threshold between the reflectance abrupt change curve of the candidate defect region in the multispectral band and the preset defect feature curve.

[0011] Generate a 3D topography reconstruction command for defects, and trigger a laser displacement sensor to collect surface elevation data based on the coordinates of candidate defect regions verified by spectroscopy.

[0012] A defect classification decision tree is established. Elevation data and multispectral features are input into the classifier, and the final defect category is output according to the combination rules of material type and deformation degree.

[0013] Preferably, the separation of the iron oxide scale-covered area and the base metal area by the difference in spectral reflectance includes:

[0014] Obtain the ratio of reflectance intensity in the near-infrared band to that in the visible light band. When the ratio exceeds the first threshold, it is marked as an iron oxide scale area.

[0015] Calculate the difference in absorptivity between the short-wave infrared band and the mid-wave infrared band, and mark the region as a base metal region when the difference is lower than the second threshold.

[0016] The material distribution map is generated by fusing multi-band labeling results, and pixels with labeling conflicts are reclassified using a neighborhood voting mechanism.

[0017] Preferably, the construction of the dynamic threshold segmentation model includes:

[0018] The grayscale histogram distribution of each region in the statistical material distribution map is used to extract the grayscale peak range of the iron oxide scale region as a benchmark reference value.

[0019] Establish a grayscale fluctuation tracking window to monitor the offset of the baseline reference value in adjacent frame images in real time;

[0020] When the offset exceeds the preset tolerance, the segmentation interval boundary is recalculated based on the average grayscale value of the most recent ten frames.

[0021] Asymmetric threshold compensation is applied to the base metal region, and a double tolerance band is set in the range where the gray value is lower than the reference value.

[0022] Preferably, the establishment of the defect morphology feature library includes:

[0023] Load the standard defect template library, which includes three levels of morphological feature descriptions of cracks, inclusions, and pores;

[0024] Perform skeleton extraction on the candidate defect region to generate a simplified contour with the same topological structure as the standard template;

[0025] Calculate the Hausdorff distance between the simplified contour and the standard templates at each level, and filter the candidate templates whose distance values ​​are in the top 20%.

[0026] Verify whether the morphological parameters of the candidate templates satisfy the aspect ratio constraint and curvature continuity condition.

[0027] Preferably, the spectral verification of the defective region includes:

[0028] Extract the reflectance values ​​of candidate defect regions in six characteristic bands to generate normalized reflectance distribution curves;

[0029] Calculate the absolute value of the slope difference between the distribution curve and the preset defect characteristic curve in the key band;

[0030] When the slope difference of more than three bands is below the allowable error, the spectral verification is deemed successful.

[0031] For candidate regions that fail validation, a second scan is initiated, and data is reacquired using a higher-precision spectrometer.

[0032] Preferably, the instruction for generating the three-dimensional topography reconstruction of defects includes:

[0033] Based on the center coordinates of the defect region that has passed spectral verification, a scanning path plan for the laser displacement sensor is generated.

[0034] The control sensor scans along the main axis of the defect in a sawtooth pattern, with the acquisition interval not exceeding one-fifth of the surface roughness;

[0035] Perform outlier filtering on the collected elevation data, removing data points that exceed three times the standard deviation;

[0036] A non-uniform rational B-spline interpolation algorithm is used to reconstruct a continuous three-dimensional surface model of the defect region.

[0037] Preferably, the establishment of the defect classification decision tree includes:

[0038] The first-level classification node is defined as material type discrimination, which distinguishes between iron oxide scale defects and base metal defects based on multispectral characteristics;

[0039] The second-level classification node is set as deformation depth discrimination, which compares the maximum value of elevation data with the standard value of thickness to generate depth level;

[0040] Configure the third-level classification node as defect cause reasoning, and combine the depth level and material type to match abnormal process parameter records;

[0041] Output the final defect category code, which includes a three-part code of defect type, severity, and possible cause.

[0042] Preferably, the definition of the first-level classification node includes:

[0043] Extract the reflectance ratio of the defect region in the characteristic band and match it with the reference spectrum in the material database;

[0044] When the short-wave infrared absorption rate exceeds the threshold, the iron oxide scale defect branch is activated.

[0045] Calculate the near-infrared band reflection intensity gradient, and activate the matrix metal branch when the gradient change rate is in the metal phase transition range.

[0046] For abnormal spectral features that cannot be matched, initiate a manual review process, suspend automatic classification, and send an alarm signal.

[0047] Preferably, setting the second-level classification node includes:

[0048] Read the maximum elevation difference in the 3D surface model and convert it into a deformation percentage relative to the standard thickness;

[0049] When the percentage exceeds a critical threshold, it is marked as a severe defect in the depth level;

[0050] Calculate the elevation change gradient of the defect edge region, and activate the emergency processing flag when the gradient exceeds the deformation rate corresponding to the material yield strength.

[0051] Generate priority handling instructions for cases that simultaneously trigger both critical defects and emergency handling flags.

[0052] Preferably, the configuration of the third-level classification node includes:

[0053] Search the historical database of continuous casting process parameters and filter records of abnormal cooling rates that match the time of defect occurrence;

[0054] Compare the current defect morphological characteristics with the typical causal patterns in historical cases to assess their similarity.

[0055] When the crystallizer vibration parameters are detected to match the periodic characteristics of the current defect, an equipment fault identifier is added to the cause code.

[0056] The output includes a complete inspection report containing defect location coordinates, classification codes, and cause predictions.

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] This image processing-based method for detecting surface defects in continuously cast billets acquires multispectral image sequences of the billet surface. By utilizing differences in spectral reflectance to separate the oxide scale-covered area from the base metal area and establishing a surface material distribution map, it can accurately obtain the distribution of different materials on the billet surface, avoiding misjudgments due to material confusion. Compared to traditional single-spectrum detection, multispectral image sequences contain richer optical information, clearly presenting the characteristic differences of different material regions. This provides more comprehensive and accurate basic data for subsequent defect detection, reducing detection bias caused by inaccurate material identification.

[0059] A dynamic threshold segmentation model is constructed to generate adaptive segmentation intervals based on the grayscale value fluctuation characteristics of different regions in the material distribution map, extracting the contour boundaries of potential defect areas. This process does not require manually setting a fixed threshold and can dynamically adjust the segmentation standard according to the actual grayscale value changes on the continuous casting billet surface. During continuous casting production, the grayscale value of the continuous casting billet surface will change due to factors such as production environment and process parameter fluctuations. The dynamic threshold segmentation model can adapt to these changes in real time, ensuring accurate extraction of the contours of potential defect areas under different operating conditions. This avoids the inaccurate segmentation problem that occurs when fixed thresholds are used when grayscale values ​​fluctuate, improving the adaptability and stability of potential defect area identification.

[0060] A defect morphology feature library is established, and the contour boundaries are matched with standard defect templates at multiple scales to filter candidate defect regions that meet geometric constraints. Multi-scale matching allows for a comprehensive comparison of the morphological differences between the contour boundaries and the standard templates. Combined with further filtering based on geometric constraints, interference regions that do not conform to the defect's geometric characteristics can be eliminated, accurately locating candidate defect regions. Compared to traditional methods relying solely on single-scale matching, this approach, combining multi-scale matching with geometric constraints, can more meticulously identify the morphological features of defects, reduce interference from non-defect regions on the detection results, and improve the accuracy of candidate defect region selection.

[0061] Spectral verification of the defect area is performed by comparing the reflectance abrupt change curves of the candidate defect area in multiple spectral bands with a preset defect feature curve based on a similarity threshold. By comparing the multi-band reflectance abrupt change curves, the authenticity of the candidate defect area can be further verified from the perspective of spectral characteristics. Different types of defects exhibit specific patterns in reflectance changes across multiple spectral bands. By comparing the similarity with the preset defect feature curve, false defects caused by factors such as surface stains and changes in illumination can be eliminated, ensuring that the selected candidate defect areas possess the spectral characteristics of genuine defects and improving the reliability of defect identification.

[0062] The system generates a 3D defect topography reconstruction command, triggering a laser displacement sensor to collect surface elevation data. Based on this elevation data, the 3D defect topography is reconstructed, providing a clear view of the defect's spatial morphological characteristics. Traditional 2D image detection can only acquire planar information about defects, failing to accurately reflect spatial parameters such as depth and height. 3D topography reconstruction overcomes this deficiency, clearly displaying the three-dimensional morphology of defects on the continuous casting billet surface. This provides richer spatial data for subsequent defect analysis, facilitating a more comprehensive understanding of the defect's actual condition.

[0063] A defect classification decision tree is established by inputting elevation data and multispectral features into a classifier. Based on a combination of material type and deformation degree rules, the final defect category is output, achieving accurate defect classification and causal correlation. By integrating multi-dimensional data, the classification decision tree can systematically analyze the material properties and deformation characteristics of defects, accurately classify defect types according to preset rules, and infer possible causes by combining material type and deformation degree. This provides a clear direction for adjusting process parameters in the production process, facilitating targeted optimization of the production flow, reducing the recurrence of similar defects, and promoting continuous improvement in continuous casting production quality. Attached Figure Description

[0064] Figure 1 This is a schematic diagram illustrating the working principle of the image processing-based continuous casting billet surface image defect detection method described in this invention.

[0065] Figure 2 A flowchart of a method for separating iron oxide scale from the base metal region;

[0066] Figure 3 A visualization of multispectral detection and analysis of surface defects in continuously cast billets;

[0067] Figure 4 A flowchart for establishing a defect morphology feature library. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Please see Figure 1This invention provides a method for detecting surface defects in continuously cast billets based on image processing. The method combines multispectral imaging technology with three-dimensional morphology analysis to achieve automated identification and classification of surface defects in continuously cast billets. The core of the method lies in integrating multi-band spectral information and elevation data, and establishing a multi-level verification mechanism to ensure the accuracy and reliability of defect detection. The implementation process involves acquiring a sequence of multispectral images of the continuously cast billet surface, separating the oxide scale-covered area from the base metal area using the spectral reflectance characteristics of different materials, and generating a surface material distribution map. Based on the material distribution map, a dynamic threshold segmentation model is constructed, adaptively adjusting the segmentation interval according to the grayscale fluctuation characteristics of different regions to extract the contour boundaries of potential defect areas. Multi-scale matching is performed using a defect morphology feature library, comparing the contour boundaries with standard defect templates to screen candidate defect areas that meet geometric constraints. Candidate areas need to undergo spectral verification, analyzing the similarity between their multispectral reflectance abrupt change curves and preset defect feature curves, and determining the verification validity through a threshold. For areas that pass verification, a laser displacement sensor is triggered to acquire surface elevation data, and three-dimensional morphology reconstruction of the defects is performed. A defect classification decision tree is established. Elevation data and multispectral features are input into the classifier, and the final defect category is output based on the combination rules of material type and deformation degree, thus completing a fully automated detection process.

[0070] Example 1: See Figure 2 The material region is separated and a dynamic threshold model is constructed through multispectral image processing. The whole process begins with the acquisition of multispectral image sequences. A high-resolution multispectral camera covering the visible light to mid-infrared band is used. The camera is installed above the continuous casting billet transport path to ensure that the imaging plane is parallel to the surface. The image acquisition frequency is dynamically adjusted according to the continuous casting billet moving speed, usually set to 10 frames per second to match the production line speed and avoid motion blur. The acquired image sequence is transmitted to the processing unit in real time for preprocessing such as flat field correction and dark current compensation to eliminate sensor noise and the influence of ambient light. When separating the iron oxide scale-covered area from the base metal area by the difference in spectral reflectance, the system extracts the reflectance intensity value of each pixel in a specific band from the multispectral image. The near-infrared band is selected in the range of 700-1100 nm, and the visible light band covers 400-700 nm. The ratio of reflectance intensity is calculated by division. The ratio result is compared with a preset first threshold of 1.5. When the ratio of a pixel exceeds the threshold, it is automatically marked as an iron oxide scale area. At the same time, the system processes the data in the short-wave infrared band of 1500-2500 nm and the mid-wave infrared band of 3000-5000 nm. The difference in absorptivity is calculated by subtraction. When the difference is lower than the second threshold of 0.1, it is marked as a base metal area. The marking process is performed pixel by pixel to generate an initial material distribution map.

[0071] Multi-band marking may generate conflicting pixels. For example, the same location may show a high ratio in near-infrared but a low difference in short-wave infrared. The system initiates a neighborhood voting mechanism to resolve the conflict. A 3×3 pixel neighborhood is defined with the conflicting pixel as the center. The number of confirmed iron oxide scale or metal markers in the neighborhood is counted. If the proportion of iron oxide scale markers exceeds 60%, the point is reclassified as iron oxide scale; otherwise, it is classified as metal. When the proportions are equal, iron oxide scale markers are retained first to enhance defect sensitivity. After voting, a final material distribution map is generated. This map is stored in the form of a two-dimensional matrix, with each element encoding the material type. Iron oxide scale is represented by a value of 1, and metal is represented by a value of 2. The map serves as the basic input for subsequent processing.

[0072] The construction of a dynamic threshold segmentation model relies on the grayscale information of the material distribution map. The system loads the map and maps it to a grayscale image sequence, statistically analyzes the grayscale histogram distribution of the iron oxide scale region, and uses a 256-level grayscale histogram to analyze the concentrated range of pixel values. For example, in a typical 8-bit image, the grayscale values ​​of the iron oxide scale region are mostly concentrated in the range of 150-200. The reference value is determined by finding the peak value of the histogram, and the median of the peak range, 175, is often taken as the reference. At the same time, the grayscale distribution of the base metal region is analyzed, and its values ​​are usually lower and more dispersed. A grayscale fluctuation tracking window is established to monitor the stability of the reference value. The window size is set to 5 consecutive frames of images, and the offset of the reference value between adjacent frames is calculated in real time. The offset is obtained by the mean difference method, that is, comparing the absolute difference between the current frame's reference value and the previous frame's value. When the offset exceeds the preset tolerance of 10, the system determines that the illumination or surface conditions have changed, triggering a recalculation of the threshold. The recalculation process is based on the data of the most recent ten frames of images, extracting the average and standard deviation of the grayscale values ​​of the iron oxide scale region in the ten frames. The boundary of the new segmentation interval is set to the average value plus or minus twice the standard deviation to ensure adaptability to environmental changes. Asymmetric threshold compensation is applied to the base metal region. Since the gray value of the metal region is generally lower than that of iron oxide scale, the compensation strategy sets a double tolerance band in the range where the gray value is lower than the reference value. For example, when the reference value is 175, the lower limit of metal region segmentation is set to 155 (reference value minus 20), and the upper limit is set to 170 (reference value minus 5). After compensation, threshold segmentation is performed to generate a binary image in which the high gray value region is the candidate defect. The contour boundary of the potential defect is extracted by the edge detection algorithm such as the Canny operator. The boundary point coordinates are stored in chain code sequence format for easy subsequent morphological matching.

[0073] The hardware configuration for multispectral image acquisition involves camera calibration and band selection. The camera uses a prism-based multispectral sensor, with each band corresponding to an independent detector. The calibration process includes whiteboard reference measurements to standardize reflectance. During acquisition, a uniform halogen light source is used for surface illumination of the continuously cast billet to avoid shadow interference. The image sequence is stored as a multi-channel data structure, with each channel corresponding to one band. The processing unit parses the channel data frame by frame, and the reflectance intensity is calculated based on pixel brightness values ​​normalized to the 0-1 range. In the spectral reflectance difference analysis, the ratio and difference calculations use floating-point operations to ensure accuracy. The threshold setting is based on historical data statistics. The first threshold of 1.5 is derived from the typical differences between iron oxide scale and metal in the near-infrared-visible light band, while the second threshold of 0.1 is based on the short-wave infrared absorption characteristics. The conflict resolution mechanism also considers edge pixel processing. For conflict points at image boundaries, the neighborhood voting uses a mirror-fill method to expand the neighborhood, ensuring the validity of the vote. After the material distribution map is generated, the system performs consistency checks, such as filtering isolated pixels and using morphological opening operations to eliminate mislabeling caused by noise.

[0074] The implementation process of constructing the dynamic threshold segmentation model begins with a deep analysis of the generated material distribution map. The system groups the pixels in the map according to the material classification results and separately analyzes the grayscale value distribution of the iron oxide scale region and the base metal region. The statistical process uses a sliding window method to traverse the entire image region. The window size is set to 32×32 pixels. The grayscale histogram of the specified material pixels in each window is calculated. The histogram is divided into 256 grayscale levels. The dominant grayscale feature of the region is determined by finding the peak interval of the histogram. For example, in a typical 8-bit grayscale image, the grayscale values ​​of the iron oxide scale region are mostly concentrated in the range of 150 to 200. The system uses the grayscale value with the highest frequency in this range as the benchmark reference value for the region. A grayscale fluctuation tracking window is established to monitor the impact of changes in lighting conditions or surface conditions in the production environment on grayscale values. The tracking window is set as a time series of 5 consecutive frames. The system compares the difference between the benchmark reference value of the iron oxide scale region in the current frame and the previous frame in real time. This comparison is achieved by calculating the absolute difference between the baseline values ​​of two adjacent frames. When the system detects that the offset exceeds the preset tolerance level of 10 gray levels, it determines that the current environmental conditions have changed significantly and the segmentation threshold needs to be recalibrated. At this time, the system will automatically retrieve the grayscale data of the iron oxide scale region in the most recent ten frames of images, calculate the arithmetic mean and standard deviation of the grayscale values ​​in these ten frames, and set the new segmentation interval boundary as the range of the mean plus or minus twice the standard deviation. This dynamic adjustment mechanism enables the model to adapt to the light fluctuations and surface condition changes on the production line.

[0075] The asymmetric threshold compensation strategy applied to the substrate metal region is based on the difference in optical properties between the metal surface and the iron oxide scale surface. The metal region typically has lower reflectivity, and its grayscale value is generally lower than that of the iron oxide scale region. The compensation operation sets a wider tolerance band in the range where the grayscale value is lower than the reference value. Specifically, a larger offset is subtracted from the reference value as the lower limit of the segmentation interval, and a smaller offset is subtracted as the upper limit. For example, when the reference value is 175, the lower limit of the metal region segmentation may be set to 155 and the upper limit to 170. This asymmetric processing can effectively avoid false detections caused by normal grayscale fluctuations on the metal surface, while ensuring the detection sensitivity for low grayscale defects. The threshold segmentation operation uses dynamically generated interval boundaries to process the entire image, generating a binary image in which high grayscale regions are marked as potential defects. Subsequently, the contour boundaries of these regions are extracted using an edge detection algorithm. Throughout the model's operation, various parameters such as tracking window size, tolerance threshold, and standard deviation multiple can be adjusted through the configuration file, enabling the system to adapt to the specific operating conditions of different continuous casting production lines. The model update frequency is synchronized with image acquisition to ensure real-time processing capabilities.

[0076] See Figure 3 In the continuous casting billet surface defect detection method based on multispectral image processing, the core data analysis and result visualization embody the integrated process of material separation, threshold segmentation, and defect identification. Specifically, multispectral band data acquisition compares the normalized reflectance of iron oxide scale and the base metal in characteristic bands (such as visible light 400-700nm and near-infrared 700-1100nm) to establish a reflectance comparison histogram. A threshold line of 0.5 is set as the quantitative benchmark for material discrimination, corresponding to the key step in the project background of separating material regions based on spectral reflectance differences. The material region separation results are presented in the form of a binary image matrix. In an 80×80 pixel grid, white pixels encode iron oxide scale regions, and black pixels encode base metal, generating a surface material distribution map as input to the dynamic threshold segmentation model. The "potential defect region" label in the image indicates the starting point for subsequent contour extraction. The near-infrared / visible reflectance ratio distribution heatmap uses a 0-2.0 grayscale gradient to encode spatial heterogeneity, with a threshold of 1.5 indicated by dashed lines. This quantifies material distribution differences to support defect spectral verification. This part is related to the comparison of spectral abrupt change curves in the defect area and the generation of adaptive thresholds in the associated projects. Parameter configuration is implemented as per Example 1, such as image acquisition frequency synchronized with the continuous casting billet movement speed, band selection covering a multispectral range, pixel-level processing to ensure detection accuracy, and the overall layout following a logical progression from data acquisition to defect identification to verify the effectiveness of the algorithm.

[0077] Example 2: See Figure 4The establishment of the defect morphology feature library and the execution of the spectral verification process are carried out in this stage. This stage takes over the candidate defect contour data generated by dynamic threshold segmentation and aims to screen high-confidence defect regions through morphological matching and spectral analysis. The construction of the defect morphology feature library begins with loading a predefined standard defect template library. The template library is stored in digital format and contains three-level morphological feature descriptions of typical defects such as cracks, inclusions, and pores. The first-level features define the macroscopic geometric parameters of the defect, such as the contour area and the aspect ratio of the circumscribed rectangle. The second-level features describe the microscopic structural properties, such as the number of extreme points of boundary curvature and the Fourier descriptor coefficients. The third-level features characterize the topological properties, such as the number of connected branches of the contour and the distribution of skeleton nodes. The system performs skeleton extraction on the candidate defect regions output by dynamic threshold segmentation. An iterative refinement algorithm is used to peel off the boundary pixels layer by layer until a central skeleton with a single pixel width is generated. During the skeleton extraction process, the topological connectivity of the original contour is preserved, and the branch points are smoothed to avoid the generation of pseudo-skeletons. The obtained simplified contours are scaled and normalized with the standard templates. The minimum circumscribed circle of the contour is calculated to achieve proportional scaling.

[0078] In the multi-scale matching stage, the Hausdorff distance between the simplified contour and the standard templates at each level is calculated. This distance metric is calculated bidirectionally, obtaining the maximum and minimum Euclidean distances from the contour point set to the template point set, and taking the maximum value of the bidirectional distance as the final matching degree index. After calculating the distances of all templates, the system sorts them and selects the candidate templates with Hausdorff distance values ​​in the top 20% to enter the verification stage. The verification process imposes strict morphological constraints. The aspect ratio constraint requires that the aspect ratio of the minimum bounding rectangle of the candidate defect region differs from that of the matching template by no more than 15%. The curvature continuity condition is verified by calculating the rate of curvature change of the contour point set. The curvature difference between adjacent contour points is calculated using the difference method, requiring that the curvature difference between consecutive points be lower than the threshold of 0.05. Candidate templates that do not simultaneously meet both constraints are excluded. The verified defect area enters the spectral verification stage. The system extracts the reflectance values ​​of the corresponding coordinate points in six characteristic bands from the multispectral image database. The characteristic bands are selected based on the defect spectral response characteristics, covering the visible light bands of 450 nm, 550 nm, and 650 nm, and the infrared bands of 850 nm, 1200 nm, and 3500 nm. The reflectance values ​​are processed by dark current correction and flat field normalization to generate a distribution curve in the range of zero to one.

[0079] The spectral verification process compares the reflectance distribution curve of the candidate defect region with a preset defect characteristic curve, which is derived from a historical defect database. Each curve defines a slope characteristic value for a key band. The slope difference calculation selects two key intervals: 550-650 nm and 1200-3500 nm. The rate of reflectance change within these intervals is calculated through linear fitting, and the absolute value of the slope of the fitted line is used as the comparison index. When the slope difference across more than three bands is less than the allowable error of 0.1%, the system determines that the spectral verification has passed. The allowable error value is dynamically adjusted based on the sensor's accuracy, typically set to 10% of the standard slope value. For candidate regions that fail verification, the system automatically initiates a secondary scanning process, triggering a high-precision spectrometer to reacquire data. The spectrometer resolution is increased to the one-nanometer level, and the sampling point density is doubled. If the secondary verification still fails, the region is marked as a false defect and excluded from the processing flow.

[0080] The entire verification process adopts a pipeline architecture, with morphological matching and spectral analysis processed in parallel. Matching results are stored in a temporary buffer awaiting spectral verification signals. The coordinates and feature parameters of the verified defect areas are packaged and transmitted to the 3D reconstruction module. The defect morphology feature library is updated and maintained using an incremental learning mechanism. When a newly discovered defect type is manually confirmed, its morphological features and spectral data are added to the template library, and the template library version management records each update. The skeleton extraction algorithm is optimized for the surface characteristics of continuously cast billets, and interpolation is used to connect fracture contours to avoid contour discontinuities caused by oxide scale detachment. Hausdorff distance calculation is accelerated using an approximation algorithm, simplifying the number of point sets through contour convex hulls, reducing the computational load by 30% while maintaining accuracy. The band selection in the spectral verification stage undergoes sensitivity analysis; six feature bands can effectively distinguish real defects from interference signals such as surface stains and water stains. The secondary scanning mechanism is equipped with timeout protection; when the spectrometer response times out, it automatically switches to a backup fiber optic spectrometer to continue acquisition. All verification data is recorded with timestamps and process parameters for subsequent defect cause analysis.

[0081] Example 3: Generating a 3D morphology reconstruction command for defects and executing 3D surface reconstruction. This process is initiated based on the coordinates of candidate defect regions that have passed spectral verification. The system transforms the defect region coordinates from the image coordinate system to the global coordinate system of the continuous casting billet surface. The origin of the coordinate system is set at the corner of the head of the continuous casting billet. The transformation uses an affine transformation matrix, and the matrix parameters are obtained through camera calibration to ensure coordinate accuracy within 0.1 mm. The scanning path planning uses the center point of the defect region as a reference. Principal component analysis is used to calculate the principal axis direction of the defect point set. The principal axis direction corresponds to the direction of the maximum eigenvector of the point covariance matrix, which is used to define the direction of the scanning trajectory. A zigzag trajectory planning generates a series of parallel scanning lines along the principal axis direction. The spacing between the scanning lines is dynamically adjusted according to the surface roughness. The roughness value comes from real-time data from the online measurement system. The acquisition spacing is set to not exceed one-fifth of the roughness value. For example, when the surface roughness Ra is ten micrometers, the spacing is controlled within two micrometers. The trajectory coverage extends five millimeters beyond the defect region boundary to prevent edge data loss. After receiving the path command, the laser displacement sensor control module drives the sensor to perform scanning. The sensor's moving speed is synchronized with the continuous casting billet transmission speed. Closed-loop control is used to ensure positioning accuracy. During the elevation data acquisition process, timestamps and coordinate information are collected for each point, and the data is temporarily stored in point cloud format.

[0082] Elevation data preprocessing includes outlier filtering. The filtering algorithm is based on statistical methods, calculating the mean and standard deviation of elevation values ​​for all collected points. The mean μ is calculated by summing and dividing by the number of points, and the standard deviation σ is obtained using the Bessel formula, which is as follows:

[0083]

[0084] The symbols have the following meanings: The standard deviation of elevation values. The total number of collection points. It is the index of the point. It is the elevation value of the k-th point. It is the average of all elevation values. The data points have elevation values ​​exceeding... ±3 Points within the specified range are marked as outliers and removed. The filtered point cloud is then smoothed using a moving average filter to reduce noise, with the window size set to five times the point spacing. A non-uniform rational B-spline interpolation algorithm is used to reconstruct the continuous 3D surface. This algorithm takes the filtered point cloud as input, parameterizes the point cloud data, and uses chord length parameterization to assign parameter values, ensuring that the parameter distribution is proportional to the point spacing. A control point sequence is calculated, with the number of control points adaptively selected based on the point cloud density, typically set to one-tenth of the total number of points to reduce computational load. Weight allocation is based on the local curvature of the point cloud; regions with higher curvature receive higher weights to preserve detailed features. A quadratic B-spline function is used as the basis function to ensure the C1 continuity of the surface. The interpolation process solves a system of linear equations to obtain the control point coordinates and weights, ultimately generating a smooth surface model.

[0085] After surface reconstruction, the system calculates defect morphology parameters such as maximum depth, volume, and surface area. Depth is obtained by comparing the elevation difference between the lowest point of the defect and the surrounding reference plane. The reference plane is obtained by least-squares fitting of the normal area surrounding the defect. Volume is calculated by the space between the integral surface and the reference plane. Surface area is obtained by summing the areas of triangular facets after triangulating the surface mesh. All morphology parameters are stored in a database for subsequent defect classification. The entire reconstruction process is automated without manual intervention, and the reconstruction results are visualized in real time for monitoring personnel. During implementation, the scanning path planning considers sensor physical limitations, such as acceleration and jitter compensation. The trajectory is optimized using a look-ahead algorithm to avoid data distortion caused by sharp turns. Environmental vibration during elevation acquisition is compensated for using inertial measurement unit data to ensure data accuracy. The NURBS algorithm implementation employs iterative optimization, smoothing the initial surface to eliminate fitting errors. The model output format is standardized to an STL file for compatibility with downstream processing systems.

[0086] Example 4: The defect classification decision tree uses a three-level node structure to accurately classify defects reconstructed from 3D morphology. Taking a specific case of continuous casting billet surface inspection as an example, a candidate defect located in the central region of the billet was marked during the spectral verification stage. Its coordinates are 15.7 meters from the head and 0.63 meters in the width direction. The decision tree processing begins with the first-level classification node, namely material type identification. The system extracts the reflectance ratio of characteristic bands from the multispectral data of the defect area. Specifically, it calculates the reflectance ratio of the short-wave infrared 1500 nm band to the near-infrared 850 nm band, which is 1.84. This value matches the standard reference spectral ratio of 1.79 for iron oxide scale in the material database. The difference is within the allowable range. At the same time, the short-wave infrared absorptivity of this area is calculated to be 0.73, which exceeds the preset threshold of 0.7. Therefore, the system activates the iron oxide scale defect branch. At the same time, the system calculated the near-infrared band reflection intensity gradient and analyzed the reflectance change curve in the 850 nm to 1100 nm band. The gradient change rate was 0.067 per nanometer. This value did not fall into the lower limit of the metal phase transition range of 0.05-0.1, which further confirmed the judgment of the iron oxide scale properties. The material identification was successfully completed without triggering the manual verification process for abnormal spectral characteristics.

[0087] The second-level classification node was then activated to determine the deformation depth. The system read the elevation data from the 3D surface model of the defect and calculated the elevation difference between the lowest point of the defect and the surrounding normal area to be 3.2 mm. The standard thickness of this section of the continuously cast billet is 200 mm, and the deformation percentage was calculated to be 1.6%. This value is lower than the set critical threshold of 5%, so the system marked it as a minor defect in the depth level and did not trigger the severe defect flag. Subsequently, the system calculated the elevation change gradient of the defect edge region, extracted the elevation values ​​of equally spaced points along the defect contour, calculated the elevation difference between adjacent points and divided by the point spacing to obtain the gradient distribution. The maximum gradient value was 0.04 mm per millimeter, which is lower than the deformation rate threshold corresponding to the material yield strength by 0.1 mm per second (equivalent to a spatial gradient of approximately 0.015 mm per millimeter in the scanning direction). Therefore, the emergency processing flag was not activated. Since this defect instance only triggered the minor defect flag, the system generated a regular processing instruction instead of a priority processing instruction. Throughout the classification process, the system accessed the real-time database to obtain related information. See Table 1, which shows the processing data records of this defect case at the key nodes of the decision tree.

[0088] Table 1: Data Records of Defect Case B-15.7-0.63 Classification Process

[0089] For uncertainties in material identification, the decision tree has a dedicated processing path. For example, when the short-wave infrared absorptivity is near the threshold of 0.7 (e.g., 0.69 to 0.71) and the near-infrared gradient change rate has both iron oxide scale and metallic characteristics, the system will mark the defect as "material pending" and transfer it to a parallel analysis thread. This thread will additionally check the emissivity characteristics in the mid-wave infrared band. If it still cannot be determined, automatic classification will be paused, and an alarm signal will be sent to the monitoring terminal. The alarm information includes the defect image, spectral curve, and coordinates, prompting the quality engineer to conduct a manual review. The system interface will highlight the spectral curve of the abnormal area and compare it with the reference curves of the standard iron oxide scale and the base metal to assist the engineer in making a judgment.

[0090] The data processing for deformation depth determination takes into account the influence of surface curvature. When calculating the gradient at the defect edge, the system first performs surface fitting on the reference plane, rather than simple plane fitting, to eliminate the interference of the inherent curvature of the continuously cast billet on the elevation difference calculation, ensuring the accuracy of deformation percentage and gradient values. Depth level classification is not a simple dichotomy; the system maintains a multi-level classification system. For example, deformation percentages below 1% are classified as "slight," 1% to 3% as "moderate," 3% to 5% as "significant," and above 5% as "severe." In this example, 1.6% deformation is classified as "moderate" severity, but in the output code, it is still uniformly categorized as "slight" to simplify the processing logic. All judgment results, intermediate calculation data, and triggered instructions are recorded in a classification log. This log is stored in association with the defect's 3D model and spectral data, providing a complete data chain for subsequent quality traceability and process improvement.

[0091] Example 5: Configuration and execution of defect cause reasoning. This stage, as the final node of the classification decision tree, aims to perform correlation analysis between the detected defect characteristics and continuous casting production process parameters. Taking a defect case numbered B-15.7-0.63 as an example, this defect has been identified as a minor defect of iron oxide scale by the preceding node. When the system starts at the third-level node, it searches the historical database of continuous casting process parameters. The database stores parameters such as crystallizer vibration frequency, cooling water flow rate, and billet pulling speed in a time series manner. The system sets the search time window to the data records within ten to five minutes before the defect detection timestamp. The search reveals a set of abnormal fluctuation records in the crystallizer vibration frequency during this time period. The frequency drops from 120 times per minute to 115 times per minute in a short period of time, and then quickly recovers.

[0092] The system then compares the morphological features of the current defect with typical causal patterns in the historical case database. The 3D morphology of defect B-15.7-0.63 shows a slight periodic distribution, with approximately equidistant pits distributed along the casting direction, spaced about 15 mm apart. A record exists in the historical case database with a defect morphology of periodic pits, tagged as "abnormal crystallizer vibration." The system calculates the similarity between the current defect morphology and this historical case, using contour curvature distribution and spatial frequency features for matching. The matching result exceeds a set threshold. Combined with the retrieved vibration parameter anomalies, the system determines that the periodic characteristics of the current defect are correlated with the crystallizer vibration frequency fluctuation record in both time and morphological features. Therefore, the equipment fault identifier "EQ_VIB" is appended to the generated defect cause code. While performing causal reasoning, the system cross-validates other process parameters. For example, it checks the water flow control data in the secondary cooling zone during the same time period and finds no significant deviations from the set values. The billet pulling speed curve also remains stable. This information serves as negative evidence, further reinforcing the judgment weight of vibration anomaly as the primary cause. For some complex cases, when multiple parameters are abnormal simultaneously, the system activates a weighted scoring mechanism, assigning weights based on the known correlation strength between the abnormal parameters and the defect type. The causal hypothesis with the highest score is adopted.

[0093] The system ultimately generates a complete inspection report containing defect location coordinates, classification codes, and causal inferences. The report uses a structured text format. For defect B-15.7-0.63, the report includes "Coordinates: (15.7m, 0.63m)," "Classification Code: OX-Minor-EQ_VIB," and "Cause Inference: Instantaneous fluctuations in the crystallizer vibration frequency caused uneven distribution of surface oxide scale, resulting in slight indentations." The report is transmitted to the manufacturing execution system via a standard interface. The system simultaneously triggers corresponding warning levels based on the identified cause type. In this example, due to the involvement of equipment fault identification, the system automatically generates a maintenance inspection work order, recommending a preventative inspection of the crystallizer vibration device. The entire causal reasoning process relies on a continuously updated historical database. Whenever a new causal case is manually confirmed, its characteristic patterns are extracted and added to the case library, continuously evolving the system's reasoning capabilities. After the report is generated, the system packages and archives all relevant data for the defect, including original images, spectral curves, 3D models, classification logs, and the final report, and links them to the product serial number of the continuously cast billet, forming a traceable quality file.

[0094] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0095] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting surface defects in continuously cast billets based on image processing, characterized in that, Includes the following steps: A multispectral image sequence of the surface of the continuously cast billet was acquired, and the area covered by iron oxide scale and the base metal area were separated by the difference in spectral reflectance to establish a surface material distribution map. A dynamic threshold segmentation model is constructed to generate adaptive segmentation intervals based on the gray value fluctuation characteristics of different regions in the material distribution map, and to extract the contour boundaries of potential defect regions. Establish a defect morphology feature library, perform multi-scale matching between the contour boundary and the standard defect template, and filter out candidate defect regions that meet the geometric constraints. Perform spectral verification of the defect region and compare the similarity threshold between the reflectance abrupt change curve of the candidate defect region in the multispectral band and the preset defect feature curve. Generate a 3D topography reconstruction command for defects, and trigger a laser displacement sensor to collect surface elevation data based on the coordinates of candidate defect regions verified by spectroscopy. A defect classification decision tree is established. Elevation data and multispectral features are input into the classifier, and the final defect category is output according to the combination rules of material type and deformation degree.

2. The method for detecting surface defects in continuously cast billets based on image processing according to claim 1, characterized in that, The method of separating the iron oxide scale-covered area from the base metal area by the difference in spectral reflectance includes: Obtain the ratio of reflectance intensity in the near-infrared band to that in the visible light band. When the ratio exceeds the first threshold, it is marked as an iron oxide scale area. Calculate the difference in absorptivity between the short-wave infrared band and the mid-wave infrared band, and mark the region as a base metal region when the difference is lower than the second threshold. The material distribution map is generated by fusing multi-band labeling results, and pixels with labeling conflicts are reclassified using a neighborhood voting mechanism.

3. The method for detecting surface defects in continuously cast billets based on image processing according to claim 2, characterized in that, The construction of the dynamic threshold segmentation model includes: The grayscale histogram distribution of each region in the statistical material distribution map is used to extract the grayscale peak range of the iron oxide scale region as a benchmark reference value. Establish a grayscale fluctuation tracking window to monitor the offset of the baseline reference value in adjacent frame images in real time; When the offset exceeds the preset tolerance, the segmentation interval boundary is recalculated based on the average grayscale value of the most recent ten frames. Asymmetric threshold compensation is applied to the base metal region, and a double tolerance band is set in the range where the gray value is lower than the reference value.

4. The method for detecting surface defects in continuously cast billets based on image processing according to claim 3, characterized in that, The establishment of the defect morphology feature library includes: Load the standard defect template library, which includes three levels of morphological feature descriptions of cracks, inclusions, and pores; Perform skeleton extraction on the candidate defect region to generate a simplified contour with the same topological structure as the standard template; Calculate the Hausdorff distance between the simplified contour and the standard templates at each level, and filter the candidate templates whose distance values ​​are in the top 20%. Verify whether the morphological parameters of the candidate templates satisfy the aspect ratio constraint and curvature continuity condition.

5. The method for detecting surface defects in continuously cast billets based on image processing according to claim 4, characterized in that, Spectral verification of the defective region includes: Extract the reflectance values ​​of candidate defect regions in six characteristic bands to generate normalized reflectance distribution curves; Calculate the absolute value of the slope difference between the distribution curve and the preset defect characteristic curve in the key band; When the slope difference of more than three bands is below the allowable error, the spectral verification is deemed successful. For candidate regions that fail validation, a second scan is initiated, and data is reacquired using a higher-precision spectrometer.

6. The method for detecting surface defects in continuously cast billets based on image processing according to claim 5, characterized in that, The command for generating the three-dimensional topography of defects includes: Based on the center coordinates of the defect region that has passed spectral verification, a scanning path plan for the laser displacement sensor is generated. The control sensor scans along the main axis of the defect in a sawtooth pattern, with the acquisition interval not exceeding one-fifth of the surface roughness; Perform outlier filtering on the collected elevation data, removing data points that exceed three times the standard deviation; A non-uniform rational B-spline interpolation algorithm is used to reconstruct a continuous three-dimensional surface model of the defect region.

7. The method for detecting surface defects of continuously cast billets based on image processing according to claim 6, characterized in that, The establishment of the defect classification decision tree includes: The first-level classification node is defined as material type discrimination, which distinguishes between iron oxide scale defects and base metal defects based on multispectral characteristics; The second-level classification node is set as deformation depth discrimination, which compares the maximum value of elevation data with the standard value of thickness to generate depth level; Configure the third-level classification node as defect cause reasoning, and combine the depth level and material type to match abnormal process parameter records; Output the final defect category code, which includes a three-part code of defect type, severity, and possible cause.

8. The method for detecting surface defects in continuously cast billets based on image processing according to claim 7, characterized in that, The definition of the first-level classification node includes: Extract the reflectance ratio of the defect region in the characteristic band and match it with the reference spectrum in the material database; When the short-wave infrared absorption rate exceeds the threshold, the iron oxide scale defect branch is activated. Calculate the near-infrared band reflection intensity gradient, and activate the matrix metal branch when the gradient change rate is in the metal phase transition range. For abnormal spectral features that cannot be matched, initiate a manual review process, suspend automatic classification, and send an alarm signal.

9. The method for detecting surface defects in continuously cast billets based on image processing according to claim 8, characterized in that, The setting of the second-level classification node includes: Read the maximum elevation difference in the 3D surface model and convert it into a deformation percentage relative to the standard thickness; When the percentage exceeds a critical threshold, it is marked as a severe defect in the depth level; Calculate the elevation change gradient of the defect edge region, and activate the emergency processing flag when the gradient exceeds the deformation rate corresponding to the material yield strength. Generate priority handling instructions for cases that simultaneously trigger both critical defects and emergency handling flags.

10. The method for detecting surface defects of continuously cast billets based on image processing according to claim 9, characterized in that, The configuration of the third-level classification node includes: Search the historical database of continuous casting process parameters and filter records of abnormal cooling rates that match the time of defect occurrence; Compare the current defect morphological characteristics with the typical causal patterns in historical cases to assess their similarity. When the crystallizer vibration parameters are detected to match the periodic characteristics of the current defect, an equipment fault identifier is added to the cause code. The output includes a complete inspection report containing defect location coordinates, classification codes, and cause predictions.

Images