A method and system for detecting the end face shape of a continuous casting slab based on binocular vision

Abstract

This invention discloses a method and system for detecting the end face shape of continuously cast slabs based on binocular vision. The method includes: constructing a binocular vision system to acquire images of the slab end face; identifying the acquired images, distinguishing between the casting head region and the slab end face region to be measured, and locating the four corner points of the slab end face; cropping the region of interest (ROI) of the end face and extracting the pixel-level contour of the slab end face from the ROI; calculating the coordinates of the four corner points in three-dimensional space according to the calibration parameters of the binocular vision system, fitting the spatial plane where the slab end face is located, and establishing a homography mapping relationship between the spatial plane and the image pixel plane; mapping the pixel-level contour extracted from the image to the spatial plane, segmenting the mapped end face contour, and performing smoothing processing, thereby calculating the shape parameters of the slab end face. This invention achieves fully automatic, high-precision, and robust online detection and quantitative analysis of the end face shape of continuously cast slabs.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and industrial automation inspection technology, and particularly relates to a method and system for detecting the end face shape of continuously cast slabs based on binocular vision. Background Technology

[0002] Continuously cast slabs are the core product of the continuous steel casting process. Their end faces serve as the benchmark for subsequent hot rolling and other processes, and their geometry and dimensional accuracy directly affect the final steel quality. However, after cutting, the slab end faces are susceptible to deformations such as edge bending, localized outward / inward expansion, and dimensional deviations due to mechanical and thermal stresses. Failure to detect and quantify these deviations in a timely and accurate manner can lead to instability in subsequent processing, resulting in decreased product quality or even production accidents. Therefore, achieving accurate online detection of the end face shape is a crucial step in improving the quality and efficiency of continuous casting production.

[0003] Currently, the determination of slab end face shape in industrial settings mainly relies on manual visual inspection. This method is inefficient, highly subjective, and struggles to obtain precise quantitative dimensions, making it prone to misjudgments and failing to meet the high precision and stability requirements of modern intelligent manufacturing. Therefore, non-contact automatic measurement technology based on computer vision has become an important development direction. However, existing visual inspection solutions face significant challenges in complex industrial scenarios. Mainstream methods are mostly based on traditional image processing algorithms (such as grayscale analysis and edge detection). These methods exhibit poor robustness and difficulty in guaranteeing measurement accuracy under the interference of high temperatures, high dust levels, uneven lighting, and changes in the target's own state in continuous casting environments. Furthermore, some systems require external specialized sensors such as hot metal detectors for triggering and positioning, resulting in complex system architectures and insufficient environmental adaptability.

[0004] Existing improvement schemes attempt to introduce more advanced technologies such as binocular vision, but when processing slab end faces against complex backgrounds, their core image processing steps (such as foreground segmentation) still heavily rely on the brightness or color features of the target, making them sensitive to lighting conditions and shooting angles. Segmentation results are poor in low-contrast or interference-prone conditions. Other methods employ pixel-by-pixel traversal for edge localization and feature analysis, resulting in low computational efficiency and weak noise resistance, leading to insufficient stability and reliability in real-time online detection scenarios.

[0005] In summary, existing technologies lack an online detection method for slab end face shape that can adapt to the harsh working conditions of continuous casting sites and achieve high robustness, high precision, and full automation. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for detecting the end face shape of continuously cast slabs based on binocular vision, comprising the following steps:

[0007] Construct a binocular vision system to synchronously acquire images of the slab end faces on the continuous casting roller conveyor;

[0008] Using a trained target detection model, the acquired images are identified to distinguish between the casting head area and the slab end face area to be tested, and the four corner points of the slab end face are located.

[0009] When the end face region of the slab is identified and the corner points are complete, the region of interest of the end face is cropped, and the pixel-level contour of the end face of the slab is extracted from the region of interest using a trained contour segmentation model.

[0010] Based on the calibration parameters of the binocular vision system and the principle of triangulation, the coordinates of the four corner points in three-dimensional space are calculated.

[0011] Based on the three-dimensional coordinates of the four corner points, the spatial plane where the end face of the slab is located is fitted, and the homography mapping relationship between the spatial plane and the image pixel plane is established.

[0012] Based on homography mapping, the pixel-level contours extracted from the image are mapped to the spatial plane to obtain end face contours with real physical dimensions.

[0013] The mapped end face contour is segmented and smoothed using a curve fitting method, and then the width, height and edge shape deviation parameters of the slab end face are calculated.

[0014] Optionally, a binocular vision system is constructed to simultaneously acquire images of the slab end faces on the continuous casting roller conveyor, specifically including:

[0015] Two industrial cameras are fixedly installed in a sealed housing above the roller conveyor to form a binocular vision system.

[0016] The cooling system continuously delivers cooling airflow into the camera's protective housing to maintain the camera's operating temperature in high-temperature environments.

[0017] Image acquisition is automatically triggered based on changes in image grayscale. The acquired images are binarized, and the area of ​​the maximum connected region is calculated. When the area of ​​the maximum connected region exceeds a preset threshold and the area change between adjacent frames exceeds a set ratio, the slab is determined to have entered the field of view, and the binocular camera is triggered to acquire images synchronously.

[0018] Optionally, before using the trained object detection model to recognize the acquired images, a vision system calibration step is also included:

[0019] A calibration board with a concentric circle array is used;

[0020] Acquire multiple images of the calibration plate from the left and right cameras in different poses;

[0021] By constructing a frontal projection model of concentric circles, the coordinates of the extracted center feature points are optimized;

[0022] Based on the optimized feature point coordinates, an iterative optimization strategy is adopted to jointly solve the internal parameters, distortion parameters, and relative pose relationship between the left and right cameras.

[0023] Optionally, the target detection model and contour segmentation model are trained based on a constructed slab end face dataset. The process of constructing the dataset includes:

[0024] After automatically acquiring images containing slabs, manual screening and classification are performed, dividing them into casting head type images and end face type images;

[0025] For images of casting heads, the casting head area is labeled; for images of end faces, the end face area and the positions of its four corner points are labeled, forming the training dataset for the target detection model.

[0026] Select a subset of samples from end-face type images and perform pixel-level fine annotation on the complete contour of the slab end face to form a training dataset for the contour segmentation model.

[0027] Optionally, the trained object detection model is used to identify the image, specifically:

[0028] The YOLO series target detection network is used to identify the end face area and the casting head area, and the bounding box of the end face area and the coordinates of the four corner points are output.

[0029] The subsequent contour extraction and measurement process is executed only when the detection target is the end face of the slab and all four corner points are successfully located.

[0030] Optionally, the pixel-level contour of the slab end face is extracted using a trained contour segmentation model, specifically:

[0031] An image segmentation network with a U-Net structure is used to process the region of interest on the cropped end face and output a binarized contour mask of the end face.

[0032] Based on the approximate location of the corner points provided by the target detection model, the nearest contour point is searched on the extracted contour to accurately locate the actual corner point pixel coordinates.

[0033] Optionally, based on the calibration parameters of the binocular vision system and the principle of triangulation, the coordinates of the four corner points in three-dimensional space are calculated, specifically as follows:

[0034] Use the left camera coordinate system as the world coordinate system;

[0035] For the matching corner pixel coordinates in the left and right images, a system of equations is established by combining the intrinsic parameters and distortion parameters of the left and right cameras, as well as the rotation matrix and translation vector of the right camera relative to the left camera.

[0036] The least squares method is used to solve the system of equations to obtain the three-dimensional spatial coordinates of each corner point in the world coordinate system.

[0037] Optionally, based on the three-dimensional coordinates of the four corner points, the spatial plane containing the end face of the slab is fitted, and a homography mapping relationship is established, specifically as follows:

[0038] Calculate the centroids of the four corner points in 3D coordinates, and calculate the offset matrix of each corner point relative to the centroids;

[0039] Perform singular value decomposition on the offset matrix, and take the left singular vector corresponding to the minimum singular value as the normal vector of the fitting plane;

[0040] Each corner point is projected onto the fitting plane to obtain the projected points on the plane;

[0041] Establish a two-dimensional coordinate system in the fitting plane with the centroid as the origin;

[0042] The two-dimensional coordinates of the projection point in the two-dimensional coordinate system are calculated. By combining these two-dimensional coordinates with the corresponding pixel coordinates in the original image, the homography transformation matrix is ​​calculated.

[0043] Optionally, the mapped end face contour is segmented and smoothed using a curve fitting method, specifically:

[0044] The end face profile mapped onto the actual size plane is segmented along the edge direction at fixed intervals.

[0045] For each segment of the contour point set, a third-order Bézier curve is used for fitting;

[0046] Based on the fitted curve, the width and height of the slab end face are calculated, and the edge shape deviation is obtained by calculating the distance from the point on the contour to its corresponding edge reference line.

[0047] On the other hand, this embodiment also provides a binocular vision-based continuous casting slab end face shape detection system for implementing any of the methods described above, including a hardware module and a software module;

[0048] The hardware module includes:

[0049] The camera system module is used to synchronously acquire images of the end face of the slab running on the continuous casting roller conveyor;

[0050] A cooling system module, connected to the camera system module, is used to deliver cooling airflow into the camera system module to ensure stable operation in high-temperature industrial environments.

[0051] The software module is deployed on an industrial control computer, and the software module includes:

[0052] The vision system calibration module is used to calibrate the parameters of the camera system module and establish a mapping relationship between image coordinates and three-dimensional space.

[0053] The slab end face contour detection module is used to receive the acquired images and, based on the mapping relationship established by the vision system calibration module and the trained deep learning model, identify the slab end face region from the images, locate its four corner points, and extract the pixel-level contour of the end face.

[0054] The slab end face contour shape measurement module is used to receive corner point and contour information, and calculate the actual width, height and edge shape deviation parameters of the slab end face according to the calibration parameters through three-dimensional reconstruction, plane fitting, homography mapping and curve fitting algorithms.

[0055] The human-computer interaction module is used to control the system's start and stop, display the process and results of image acquisition, contour extraction, and parameter calculation in real time, and manage and export measurement data.

[0056] Compared with the prior art, the present invention has the following advantages and technical effects:

[0057] This invention utilizes a binocular vision system to synchronously acquire images of the end face of a running slab, and combines this with deep learning-based intelligent recognition technology to achieve automated, high-precision positioning and extraction of the slab end face region and its key corner points. This technique enables the system to effectively distinguish between the casting head and the end face to be tested, and accurately segment the contour in complex industrial environments. It significantly reduces reliance on manual experience and fixed lighting conditions, and substantially improves the automation level of the inspection and its environmental robustness under harsh conditions such as high temperatures and dust.

[0058] This invention, based on high-precision vision system calibration parameters and triangulation principles, converts two-dimensional image information into three-dimensional spatial coordinates of the corner points of the slab end face. Through spatial plane fitting and homography mapping techniques, it accurately maps the pixel-level contours in the image to a spatial plane with true physical dimensions. This series of techniques achieves accurate reconstruction from two-dimensional images to three-dimensional dimensions, freeing the measurement results from the constraints of pixel units and obtaining high-precision, true-dimensional data that can be directly used for process guidance.

[0059] This invention performs piecewise Bézier curve fitting on the mapped real contour. This process not only smooths out measurement noise, making the contour curve smoother and more reasonable, but also provides a reliable data foundation for quantitative analysis of edge shape deviations. The method ultimately outputs the width, height, and quantified edge shape deviation parameters of the slab end face, realizing a transformation from qualitative judgment to precise quantitative analysis of end face shape. This provides stable and reliable data support for quality control, process parameter optimization, and prevention of edge warping and blockage problems in hot rolling processes during continuous casting production, thus demonstrating significant comprehensive value in improving product quality consistency and production efficiency. Attached Figure Description

[0060] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0061] Figure 1 This is a flowchart illustrating the binocular vision-based end face shape detection method for continuously cast slabs provided in an embodiment of the present invention.

[0062] Figure 2 A schematic diagram of the structural composition of the binocular vision-based continuous casting slab end face shape detection system provided in an embodiment of the present invention;

[0063] Figure 3 This is a schematic diagram of the concentric circle orthographic projection model used for vision system calibration in an embodiment of the present invention;

[0064] Figure 4 This is a flowchart of the end face contour detection process in an embodiment of the present invention;

[0065] Figure 5 This is a schematic diagram of the model for measuring the end face contour of a continuously cast slab in an embodiment of the present invention. Detailed Implementation

[0066] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0067] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0068] Example 1

[0069] This embodiment provides a method and system for detecting the end face shape of continuously cast slabs based on binocular vision. Specifically, the detection method includes the following steps:

[0070] Construct a binocular vision system to synchronously acquire images of the slab end faces on the continuous casting roller conveyor;

[0071] Using a trained target detection model, the acquired images are identified to distinguish between the casting head area and the slab end face area to be tested, and the four corner points of the slab end face are located.

[0072] When the end face region of the slab is identified and the corner points are complete, the region of interest of the end face is cropped, and the pixel-level contour of the end face of the slab is extracted from the region of interest using a trained contour segmentation model.

[0073] Based on the calibration parameters of the binocular vision system and the principle of triangulation, the coordinates of the four corner points in three-dimensional space are calculated.

[0074] Based on the three-dimensional coordinates of the four corner points, the spatial plane where the end face of the slab is located is fitted, and the homography mapping relationship between the spatial plane and the image pixel plane is established.

[0075] Based on homography mapping, the pixel-level contours extracted from the image are mapped to the spatial plane to obtain end face contours with real physical dimensions.

[0076] The mapped end face contour is segmented and smoothed using a curve fitting method, and then the width, height and edge shape deviation parameters of the slab end face are calculated.

[0077] The purpose of this embodiment is to achieve accurate measurement and determination of the end face profile of slabs transported on the rolling mill. It aims to solve the problems of low accuracy and inefficiency in traditional manual observation and measurement methods based on simple image processing, providing reliable data support for slab end face quality control, process optimization, and production stability.

[0078] This embodiment primarily employs a binocular vision system composed of two industrial cameras as vision sensors to acquire image information of the slab end face on the rolling mill. It then combines this with a deep learning-based image processing algorithm to achieve ROI region identification and contour measurement of the slab end face. Through processing steps such as end face position detection, region extraction, and contour segmentation, the end face contour and actual size data are obtained. The measurement results are then statistically recorded, enabling precise measurement and analysis of the slab end face shape.

[0079] The method for detecting the end face shape of continuously cast slabs proposed in this embodiment includes the following main steps:

[0080] Industrial Site Hardware System Setup: The hardware system in this embodiment mainly consists of a binocular vision system encapsulated in a metal casing, an industrial control computer, and a matching air cooling device. The binocular vision system is used to acquire images of the end face of the continuously cast slab running on the roller bed, providing basic data for subsequent end face contour measurement; the industrial control computer is used to complete image processing and execute measurement algorithms. Each camera is equipped with an independent protective housing, which adopts a sealed and heat-insulating structure design to isolate the high temperature and heat radiation effects of the production site. The air cooling device continuously delivers cooling airflow into the protective housing to remove the heat generated during system operation, thereby ensuring the stable and reliable operation of the binocular vision system in a high-temperature industrial environment.

[0081] Vision system parameter calibration: A high-precision camera calibration method based on a concentric circle calibration plate was used for calibration. This method constructs a concentric circle orthographic projection model, transforming the concentric circle imaging under non-orthographic conditions to an orthographic projection plane for feature extraction, thereby reducing the circle center offset error and improving the feature point localization accuracy. Based on this, an iterative optimization strategy is employed to jointly optimize the intrinsic parameters, distortion parameters, and extrinsic parameters of the monocular camera and the binocular camera, thus completing the camera calibration and improving the overall accuracy and robustness of the binocular measurement system.

[0082] Construction of Slab End Face Dataset: To support the training and deployment of the end face region target detection model and the end face contour segmentation model, this embodiment constructs a continuous casting slab end face dataset. The system automatically determines the presence of a continuous casting slab in the image by performing binarization threshold change analysis on the roller conveyor images, and automatically acquires images containing the slab at preset time intervals when trigger conditions are met. After acquisition, the images are manually screened and classified, dividing the slab samples into casting head type and end face type. Casting head type slabs are not included in the measurement, while end face type slabs are used as the measurement object. Based on this, two types of datasets are constructed: one is a target detection dataset, which labels the casting head face of the casting head type slab and the end face region and its four corner points of the end face type slab; the other is an end face contour segmentation dataset, which only labels the end face contour of the end face type slab at the pixel level for subsequent end face shape measurement.

[0083] Target detection and contour segmentation extraction in the end face region: Based on the constructed dataset, the target detection model and the contour segmentation model for the end face region are trained and deployed. During system operation, the target detection model is first used to distinguish between the casting head region and the slab end face region in the image. Subsequent measurement processes are only performed on the detected slab end faces, and the completeness and usability of the four corner points of the end face are determined. The system analyzes the pixel position changes of the end face center point in the image to determine the end face running position and count the number of slabs. For the detected end face region, its region of interest is further cropped, and the end face contour is extracted using the segmentation model to obtain contour data for subsequent 3D measurement.

[0084] End Face Shape Measurement: After the end face area is detected and its contour extracted, the system sends the contour data and corner point information to the background measurement thread for processing. The background thread initiates the end face shape measurement process after completing binocular data synchronization within a set time window. The system first calculates the 3D coordinates of the four corner points of the end face and performs plane fitting based on these corner points. Then, it performs homography transformation on the end face contour to obtain the true 3D dimensions of the end face contour. Based on this, the end face contour is segmented and fitted using Bézier curves to obtain the final end face shape measurement result.

[0085] Human-Computer Interaction and Data Display: This embodiment designs a human-computer interaction and data display interface for real-time monitoring of system operation status and digital management of measurement results. During the measurement process, the interface displays images of the continuously cast slab passing through the rolling mill in real time, and simultaneously overlays the end face contour curve, corner point positioning information, and the slab width, height, and edge deformation values ​​calculated in real time, achieving an intuitive display of measurement results. At the same time, the system provides a historical data management function, allowing users to query past measurement records and export selected measurement results, contour graphics, and judgment information as PDF reports for production data archiving and quality traceability.

[0086] As a specific implementation method, such as Figure 1As shown, the system designed in this embodiment consists of three main parts: a binocular vision system, a software control system, and a cooling system. The binocular vision system comprises two industrial cameras, each mounted in an independent protective housing. Temperature and humidity sensors are also installed inside the housings to monitor the cameras' operating status in real time. These two protective housings are fixed to a rectangular, sealed metal shell with screws, ensuring precise fixation and positional stability of the binocular system. The rectangular metal shell is covered with heat-insulating material and is fixedly positioned above the continuous casting roll, responsible for acquiring images of the end faces of the continuously cast slabs. The cooling system is connected to the camera protective housings, continuously supplying cooling airflow to ensure stable operation of the vision system in the high-heat-radiation industrial environment. The industrial control computer (i.e., the software control system) is located in the main control room and is connected to the binocular vision system via a high-speed communication line. It receives image data and centrally executes core algorithms such as image processing, 3D measurement, and shape determination, achieving integrated control of the entire system.

[0087] The overall flow of the continuous casting slab end face shape detection system designed in this embodiment is as follows: Figure 2 As shown, the system is mainly divided into hardware and software modules. The hardware module primarily includes: a vision system controlled by an industrial control computer to synchronously acquire image data of the end face of continuously cast slabs; reading data from temperature and humidity sensors to monitor camera operation; and a cooling system that continuously operates to ensure the stability of the equipment in high-temperature environments. The software module's workflow is divided into a preparation phase and an actual measurement phase. The preparation phase mainly involves calibrating the vision system to obtain the camera's intrinsic and extrinsic parameters, constructing the slab end face dataset, and training the relevant deep learning task network model, providing foundational support for subsequent measurement tasks. The actual measurement phase is the core operation of the system, and its specific steps include: inputting the acquired images; sequentially locating the end face region through the deep learning end face detection module and accurately extracting the end face contour through the end face segmentation module; then, combining the contour and corner point data from the left and right camera images, calculating the three-dimensional coordinates of the end face corner points, fitting the end face plane, and obtaining the final shape measurement result through Bézier curve fitting; finally, the results are displayed and saved in real time through the human-computer interaction module, and historical record queries and PDF report export functions are supported.

[0088] The steps of the above scheme will be explained in detail below:

[0089] Hardware modules:

[0090] (1) Camera system module:

[0091] This embodiment employs two high-resolution monochrome industrial cameras (5120×5120 resolution) to construct a binocular stereo vision system. The system is connected to an industrial control computer via a USB 3.0 fiber optic cable. Fiber optic transmission offers advantages such as high bandwidth and low latency, meeting the requirements for high-speed and stable transmission of high-resolution images. On the software side, a C++ control program is developed based on the camera's native SDK. This program allows for flexible adjustment of key imaging parameters, such as exposure time and gain, and displays real-time scene information through a human-machine interface, providing reliable raw image data for subsequent image processing and end-face contour measurement.

[0092] Considering the dusty and complex environmental conditions in industrial settings, each industrial camera is installed in an independent double-layered protective housing: the camera is placed in the inner layer, while the outer layer allows for the introduction of condensate or gas to actively cool the equipment and effectively isolate it from external dust, thus promoting long-term stable operation. Temperature and humidity sensors are installed inside the protective housing, transmitting data to an industrial control computer via serial communication for real-time monitoring of the camera's operating status. A desiccant is also included to absorb internal moisture, preventing lens fogging and ensuring image quality. To ensure the long-term stability of the relative positional relationship between the binocular cameras, the two protective housings are bolted to a custom-machined, sealed cuboid aluminum alloy structure, structurally ensuring the consistency of the baseline distance and viewing angle of the binocular system. The overall device is compact and rigid, facilitating on-site installation and long-term operation. It is ultimately fixedly deployed at a designated location above the continuous casting roll runner for continuous and stable image acquisition of the end faces of continuously cast slabs running on the roll runner during production.

[0093] (2) Cooling system module:

[0094] The industrial camera used in this embodiment operates normally within a temperature range of -20 to 70°C. However, in production environments, the surface temperature of continuously cast slabs often reaches 200 to 300°C, accompanied by significant heat radiation and convective heat transfer. This can rapidly cause the camera and its electronic components to overheat, exceeding their safe operating range and affecting image quality or even damaging the equipment. Therefore, effective thermal isolation and forced cooling of the measurement system are crucial for ensuring its stable operation. Although the camera system module has already implemented certain protective measures, a cooling system module is designed in this embodiment to guarantee the normal operation of the system.

[0095] The system mainly consists of an air compressor, an air cooler, and supporting air ducts. Its core working principle is as follows: First, the air compressor compresses ambient air; then, the compressed air is sent to the air cooler for efficient heat exchange, generating a stable, temperature-controlled, low-temperature cooling airflow. This cooling airflow is continuously and evenly injected into the internal space of the camera's protective housing through the air ducts. This design creates a continuous and stable gas flow inside the housing, effectively removing heat generated during system operation and heat infiltrating from the external environment, thus ensuring that the temperature inside the protective housing is always maintained within a safe range far below the camera's maximum operating limits. Combined with the sealed and heat-insulating structure of the protective housing, the system successfully isolates the effects of high-temperature airflow and heat radiation from the outside while ensuring that image acquisition quality is not affected, improving the system's cooling efficiency and environmental adaptability, and providing a reliable guarantee for the continuous operation of the measurement system under high-temperature and heavy-load conditions.

[0096] Software modules:

[0097] (1) Vision system calibration module:

[0098] In a binocular vision measurement system, the accuracy of camera parameters directly affects the precision and stability of the slab end face contour shape and size calculation results. Camera parameters mainly include the intrinsic parameters of each camera, distortion parameters, and the relative pose relationship between the binocular cameras. These parameters collectively determine the mapping accuracy between image pixel coordinates and the end face spatial plane, which is the foundation for realizing end face size measurement and shape analysis.

[0099] Based on the above requirements, this embodiment integrates a binocular vision calibration module for industrial field applications into the software system, which is used to complete the high-precision solution and optimization of binocular camera system parameters. After completing the initial estimation of binocular camera system parameters, this module introduces a frontal projection modeling method based on a concentric circle calibration plate, and combines iterative optimization strategies to jointly optimize the camera's intrinsic parameters, distortion parameters, and binocular extrinsic parameters, thereby obtaining high-precision camera parameters.

[0100] 1) Binocular camera parameter pre-calibration:

[0101] This invention adopts The calibration is performed using a concentric circle array calibration board. Compared with other calibration boards such as checkerboard, the concentric circle has stronger anti-noise and anti-distortion capabilities. In the feature point extraction process, the calibration board image is first subjected to edge detection, and the extracted edges are curve fitted to obtain ellipse parameters. Then, the concentric circle matching is completed according to the ellipse parameters, and finally the average value of the coordinates of the inner circle and the outer circle center is taken as the initial coordinates of the corresponding feature point.

[0102] Establish a monocular camera imaging model. Let the calibration plate coordinate system be... The coordinates of any feature point on the calibration plate in this coordinate system are: The left camera coordinate system is The right camera coordinate system is The pixel coordinates of this feature point on the left and right camera pixel planes are respectively Where u and v are in pixels, and superscript... Let represent the left and right cameras, respectively. According to the pinhole camera model, the left camera satisfies:

[0103] ;

[0104] The right camera satisfies:

[0105] ;

[0106] in, It is a homogeneous scaling factor; The intrinsic parameter matrices for the left and right cameras; It is an extrinsic parameter matrix. For rotation matrix, It is a translation vector. This represents the pose transformation from the calibration board coordinate system to the left camera coordinate system. This represents the pose transformation from the calibration board coordinate system to the right camera coordinate system. Intrinsic parameter matrix. Written as:

[0107] ;

[0108] in, Focal length is measured in pixels. The coordinates of the principal point (in pixels).

[0109] Considering the effects of lens distortion, radial and tangential distortion are used for distortion modeling. This is based on pixel coordinates. Calculate normalized image plane coordinates :

[0110] ;

[0111] make distorted normalized coordinates Represented as:

[0112] ;

[0113] in, The radial distortion coefficient is... The tangential distortion coefficient can be derived from... The distorted pixel coordinates are obtained. Based on the above distortion model, the distortion of the extracted feature point pixel coordinates can be corrected to obtain a set of distorted pixel coordinates.

[0114] Based on the aforementioned imaging and distortion model, multiple images of the calibration board in different poses were acquired, and the three-dimensional coordinates of the feature points of the calibration board were used. Its corresponding pixel coordinates in the image Zhang Zhengyou's calibration method was used to perform single-target calibration on the left and right cameras respectively, and the intrinsic parameter matrices of the left and right cameras were obtained. Distortion parameters And the calibration plate pose parameters corresponding to each calibration image. .

[0115] After completing the single-target calibration of the left and right cameras, the initial values ​​of the extrinsic parameters of the binocular system are further solved. A set of calibration images of the left and right cameras acquired at the same time are selected, and the pose parameters of the same calibration plate in the coordinate system of the left and right cameras are used. Calculate the extrinsic parameters of the binocular system This invention defines the extrinsic parameter as the transformation from the left camera coordinate system to the right camera coordinate system, which is:

[0116] ;

[0117] This allows us to obtain the initial values ​​of the extrinsic parameters of the binocular system, thus completing the pre-calibration of the binocular camera system.

[0118] 2) Iterative optimization of camera calibration parameters based on concentric circle orthographic projection model:

[0119] Because circular markers can cause eccentricity errors during feature extraction due to perspective projection and other reasons, resulting in inaccurate center coordinate extraction, this can affect the final calibration accuracy. To eliminate this effect as much as possible and make the camera calibration results more accurate, a concentric circle orthographic projection model is used to iteratively optimize the calibration results.

[0120] Considering that the concentric circle eccentricity error is caused by the non-orthographic projection of the image, the non-orthographic projection plane can be transformed into an orthographic projection plane to reduce the impact of the eccentricity error. The orthographic projection plane refers to an equivalent observation plane constructed that is parallel to the calibration board plane and whose viewing direction is consistent with the normal of the plane, so that the calibration board presents an orthographic effect without perspective distortion from this perspective. By establishing a concentric circle orthographic projection model, the feature points of the center of the standard circle are extracted on the orthographic projection plane, and then back-projected onto the original plane, thereby solving for the pixel coordinate values ​​of the center of the concentric circle in the real image.

[0121] Concentric circle orthographic projection model as follows Figure 3 As shown, let the geometric center of the concentric circles be... Consider two points M and N on the inner circle and two points A and B on the outer circle along a certain diameter of the concentric circles. Due to perspective projection, the corresponding image points of these points in the image are m, n, a, and b, respectively. Based on the image points m, n, a, and b, the pixel center of the inner circle can be calculated. The pixel center of the outer circle The actual projected position of the center of the circle in the image is denoted as . Under non-orthogonal imaging conditions , and They do not overlap, thus causing eccentricity error. After applying an orthographic correction transformation H to the model, as shown... Figure 3 As shown on the right, the calibrated calibration board plane and the camera imaging plane are in a frontal relationship, and the extracted pixel center point is obtained at this time. That is, the center of the actual concentric circles. Coordinates in the image.

[0122] Let the projection of a feature point on a non-orthographic projection plane on the calibration plate be the feature point. The projection onto the frontal projection plane is the feature point. The following relationship exists between them:

[0123] ;

[0124] In the above formula, H is the homography transformation matrix from the non-orthographic projection plane to the actual projection plane. Let the plane parameters of the plane containing the calibration plate in the image in the camera coordinate system be... ,in Let be the normal vector of the plane. Let be the distance from the origin of the camera coordinate system to this plane. Since pre-calibration has been performed, the rotation and translation matrix transforming the calibration plate coordinate system to the camera coordinate system can be easily obtained. And then the plane parameters are obtained. The formula for calculating the homography matrix H is as follows (where A is the intrinsic parameter matrix of the corresponding camera):

[0125] ;

[0126] The iterative optimization camera calibration process is as follows: First, distortion correction is performed on the calibration images acquired by the left and right cameras respectively. Then, for each calibration image, its orthographic homography matrix H is calculated under the corresponding camera imaging model, and this matrix is ​​used to map the image to the orthographic viewpoint. The center feature points of concentric circles are extracted under the orthographic viewpoint. Next, the coordinates of the extracted center feature points are mapped back to the original image coordinate system through the inverse transformation of the homography matrix, and the distortion model is reintroduced to restore their coordinates. Based on the updated feature points, camera calibration is re-performed to correct the calibration parameters. This process is iterated continuously until the final calculated reprojection error converges and meets the preset allowable error range. (Reprojection error) The calculation formula is as follows:

[0127] ;

[0128] In the above formula, The threshold is set to 0.08 when... Stop iteration when the set threshold is reached; , , Let the reprojection errors of the left and right monocular cameras and the stereo camera be respectively, and let the observed pixel coordinates of the j-th feature point in the left and right cameras be respectively. The corresponding feature points calculated from the current calibration parameters are reprojected into the left and right cameras with the following pixel coordinates: Then we have:

[0129] ;

[0130] ;

[0131] ;

[0132] in, This represents the number of feature points used in the calculation by the left and right cameras. The number of matching point pairs participating in the binocular calculation.

[0133] (2) Slab end face contour detection module:

[0134] The slab end face contour detection module is a key component in measuring the shape and dimensions of the continuous casting slab end face. Its main function is to accurately locate the slab end face region from complex industrial site images and extract the end face contour in detail. Considering the complex background, significant noise interference, and continuous slab operation in continuous casting sites, this embodiment combines deep learning object detection and image segmentation methods to construct an integrated processing flow for end face detection and contour extraction.

[0135] 1) Dataset creation and annotation:

[0136] Because continuous casting production lines operate 24 / 7, the time intervals between adjacent slabs on the roll runner are not fixed. Using fixed time intervals for image acquisition would generate a large number of invalid images that do not contain slabs, increasing the burden of subsequent manual screening and processing. Therefore, this embodiment employs a binarization detection method based on image grayscale changes to automate the acquisition of slab end faces. Specifically, when a slab enters the field of view, its surface temperature is high, causing a significant change in image brightness. The system sets a grayscale threshold of 128 and an area threshold of 60,000. After binarizing the acquired image, the system extracts the largest connected region as a candidate target. When the area of ​​this region exceeds the set threshold, it determines that a slab end face exists in the current image, thus triggering image acquisition. Furthermore, to avoid repeated acquisition of slab images at the same location, the system introduces an area change constraint: the next data acquisition operation is only performed when the area of ​​the largest connected region in two adjacent frames changes by more than 10% compared to the previous frame.

[0137] Based on the above strategy, the system collected over 100,000 original images, from which 4,660 high-quality images were manually selected. Of these, 4,500 were slab end-face images and 160 were casting head images, used to construct the end-face detection dataset. Given the high accuracy and cost of annotation for end-face contour segmentation, 1,500 samples were further selected from the end-face images for fine annotation, forming the end-face contour segmentation dataset. The LabelMe tool was used to construct the dataset during the annotation process. For the end-face detection dataset, the casting head face, slab end-face region, and four corner points were annotated with bounding boxes. Corner points were recorded as point labels, but during training, square bounding boxes with 200-pixel sides were constructed centered on these points to adapt to the target detection network format. For the end-face contour segmentation dataset, the complete slab end-face contour was annotated using polygon connections. After annotation, the end-face ROIs were cropped and uniformly scaled to 640×640 pixels, and the annotated data was transformed accordingly. Ultimately, all datasets were divided into training, validation, and test sets in a 7:2:1 ratio.

[0138] 2) Network training and deployment:

[0139] For end-face region detection tasks, given the high real-time requirements in industrial settings, this embodiment employs the YOLO11 object detection network for training. The YOLO11 model boasts a lightweight structure, fast inference speed, and mature engineering capabilities, making it suitable for practical production deployment. During training, the number of training epochs is set to 200, and the optimal model is saved for each epoch. After training, the model's accuracy and recall can be evaluated on the test set. For model deployment, the weight files (.pt) obtained from Python training are converted to ONNX format, and then the model is loaded into C++ using TensorRT, enabling high-speed inference and real-time applications in industrial settings.

[0140] For the end-face contour segmentation task, this embodiment uses the U-Net network structure for training. U-Net has an encoder-decoder structure, which can effectively capture multi-scale features and is suitable for refined extraction of end-face contours. During training, the batch size is set to 4, and the number of training epochs is 20,000. During model deployment, the inference function of the segmentation model is invoked via a C++ interface calling a Python interface, and the output contour data is used for subsequent processing.

[0141] 3) End face detection and contour extraction:

[0142] The end face contour detection process is as follows Figure 4 As shown in the diagram, firstly, the image is analyzed using an object detection model to distinguish the slab end face from the casting head area. Only slabs detected as end faces with all four corners intact and usable are processed further. Then, the end face ROI region is cropped, and the slab end face contour is extracted using a segmentation model. After contour extraction, combined with the corner point positions, the contour points closest to the corner points are searched on the contour to determine the actual coordinates of the four corner points of the slab end face. This completes the accurate extraction of the end face contour and key corner points, providing reliable basic data for subsequent dimensional measurements.

[0143] (3) Slab end face contour shape measurement module:

[0144] After obtaining the slab end face contour and corner point data from the images simultaneously acquired by the left and right cameras, the shape and dimensions of the end face are measured. The entire process mainly includes three steps: First, the three-dimensional spatial coordinates of the end face edge contour points and the four corner points are calculated using the triangulation principle; second, the actual spatial plane where the slab end face is located is fitted based on the three-dimensional coordinates of these four corner points, and the homography transformation matrix between the image coordinate system and this actual plane is calculated, directly mapping the contour pixels in the image to a planar contour with real physical dimensions; finally, the mapped contour points are smoothly fitted using Bézier curves, thereby calculating the actual height, width, and edge shape deviation of the end face, such as... Figure 5 As shown.

[0145] The specific steps are as follows:

[0146] 1) Solving for the three-dimensional coordinates of the end face corner points:

[0147] Let the slab end face in the left and right camera images be the first corner points The matching pixel coordinates are respectively ,in These are pixel coordinates (in pixels), superscript. 'x' and 'y' represent the left and right cameras, respectively. It should be noted that the pixel coordinates have already been distorted using the distortion coefficients obtained from the aforementioned camera calibration; therefore, distortion parameters will not be introduced again in the following imaging relationships, and subsequent calculations will be based on the distorted pixel coordinates.

[0148] left camera coordinate system As a reference world coordinate system for binocular vision measurement systems Then the coordinates of any corner point in this reference world coordinate system are: It is easy to know:

[0149] ;

[0150] in, Scale factor; The projection matrices for the left and right cameras are defined as follows:

[0151] ;

[0152] in, The intrinsic parameter matrices of the left and right cameras are obtained from the aforementioned calibration steps. It is an extrinsic parameter matrix. For rotation matrix, This is the translation vector. Since the reference coordinate system is taken as the left camera coordinate system, the extrinsic parameters of the left camera are... ( (representing the identity matrix), the extrinsic parameters of the right camera are... , Let the extrinsic parameters of the binocular camera obtained from the aforementioned camera calibration be... ,in (or ) represents the i-th row and i-th column of the projection matrix. Given column elements i = 1, 2, 3, j = 1, 2, 3, 4. After eliminating the homogeneous scaling factor, we can obtain the following information: The system of linear equations:

[0153] ;

[0154] ;

[0155] ;

[0156] ;

[0157] Rewrite the above system of equations in matrix form , where x= The three-dimensional coordinates of the nth corner point are solved using the least squares method:

[0158] ;

[0159] The matching point set of the end face corner points of the continuously cast slab can be obtained using the above formula. 3D coordinates .

[0160] 2) End face plane fitting and contour homography mapping:

[0161] The three-dimensional coordinates of the four corner points of the slab end face are obtained. Next, the end face plane is fitted. First, the center of the three-dimensional coordinates of the corner points is calculated, and then the three-dimensional coordinates of each corner point are normalized.

[0162] ;

[0163] ;

[0164] In the above formula X is the mean center (centroid) of the set of three-dimensional coordinate points of the four corner points on the end face of the slab, and X is the centered coordinate matrix of the four corner points relative to the centroid.

[0165] Singular value decomposition of X in the above formula is as follows:

[0166] ;

[0167] In the above formula, U is the left singular vector matrix, which contains the row direction characteristics of the data, and V is the right singular vector matrix, which contains the column direction characteristics of the data. It is a singular value matrix, with the singular values ​​arranged in descending order along the diagonal. The eigenvector corresponding to the smallest singular value, i.e., the 3rd column u3 of matrix U, is selected as the normal vector of the fitting plane. The direction pointed to by this vector is the direction in which the plane fitting residual of the three-dimensional point set at the end face corner is minimized.

[0168] Then each corner point Projecting the coordinates of the corner points onto the fitting plane yields the coordinates of the corner points on the fitting plane. The projection transformation formula is as follows:

[0169] ;

[0170] To facilitate subsequent curve fitting and dimension calculations, a local two-dimensional coordinate system needs to be constructed within the fitting plane. The origin is set as the centroid. x-axis direction and the y-axis direction The calculation formula is as follows:

[0171] ;

[0172] ;

[0173] In the above formula, i and j represent two corner points on any one of the longer sides of the end face. If If the direction is inconsistent with the actual edge direction, the symbol needs to be flipped based on the point sequence.

[0174] Finally, the two-dimensional coordinates of the four corner points in this local coordinate system are:

[0175] ;

[0176] This transforms a three-dimensional point set into a set of two-dimensional point sets. Then, select one end-face contour image from the left and right images, calculate the homography matrix using the correspondence between its pixel corner points and the actual end-face plane, and map the end-face contour in the image to the actual size planar space through homography transformation.

[0177] 3) End face contour modeling and shape determination based on Bézier curves:

[0178] After obtaining the actual-sized planar end face contour, a quadrilateral formed by connecting the four corner points is used as the geometric reference for the slab end face. The width and height of the end face are obtained by calculating the distance between adjacent corner points. Subsequently, the end face contour is divided into segments of 20 pixels each along the edge direction, and a third-order Bézier curve is used to fit each segment. This method effectively smooths the contour curve and suppresses the influence of local measurement noise on the fitting results, thus obtaining a more accurate and robust end face edge shape. The Bézier curve calculation formula is as follows:

[0179] ;

[0180] in, For combinations, Let n be the control point and n be the order of the Bézier curve. In this embodiment, a third-order Bézier curve is used to model the end face profile. The control points are obtained by parametric processing of the segmented profile points and calculation using a curve fitting method.

[0181] After completing the curve fitting, the deformation of each segment contour point is calculated using the following formula, thereby achieving a quantitative analysis of the shape deviation of the slab end face.

[0182] .

[0183] (4) Human-computer interaction module:

[0184] This embodiment designs a Qt-based human-computer interaction module for the operation control and result visualization of a binocular vision continuous casting slab end face contour detection system. The interface integrates measurement control, status display, and result presentation functions. A device status indicator area is located at the bottom of the interface to display the connection and working status of each camera and related equipment. Users can start the system by clicking the "Start Measurement" button to automatically complete the image acquisition, end face contour detection, and 3D measurement process. Relevant parameters and detection results are displayed in real time during the measurement process. The "Stop Measurement" button ends the current measurement task. Simultaneously, the upper right corner of the interface displays the system running time, camera temperature and humidity information, and exposure parameters for easy monitoring of the system's operating status. (It should be noted that a color camera display window and a line laser device status indicator are reserved in the interface. These devices are currently only used for future system function expansion and on-site auxiliary display and are not currently involved in the core algorithm processes such as end face contour detection and shape measurement.)

[0185] After the measurement task is completed, users can manage and output the test results through the data export interface. The system supports filtering historical measurement data by year, month, and day, and generates corresponding test result reports. The test results can be exported as PDF documents, including visualization results of the end face contour, key dimensional parameters, and measurement time. The system categorizes and stores reports according to the test date, improving the manageability and traceability of the test results.

[0186] Compared with existing technologies, the beneficial effects achieved by this embodiment are:

[0187] This embodiment proposes a binocular vision-based method and system for detecting the end face shape of continuously cast slabs, which effectively improves the low accuracy of relying on manual visual judgment of end face shape during online slab production. This system, by constructing a real-time, non-contact three-dimensional vision measurement method, can accurately acquire the shape and size information of the slab end face during operation. This accurate and reliable measurement result has significant comprehensive value for continuous casting production: by timely identifying slabs with abnormal end faces, it reduces the risk of hot-rolled edge defects, strip breakage, and other production accidents that may be caused by misjudgment; simultaneously, accurately identifying slabs with good shapes can reduce unnecessary processing and improve hot-feeding efficiency and economic benefits. Furthermore, the system's accurate measurement of end face width, especially the lower surface width, helps prevent edge warping and blockage problems during hot rolling, providing crucial data support for process adjustments and upstream and downstream production.

[0188] Compared to existing methods using traditional image processing techniques, this embodiment offers significant advantages in robustness and measurement accuracy in complex industrial environments. At the algorithm level, this system combines deep learning-based object detection and image segmentation techniques to achieve intelligent differentiation of the casting head and end face regions in images, as well as precise extraction of the end face contour. This approach effectively overcomes the shortcomings of traditional methods, which are susceptible to interference from changes in lighting, image noise, and the slab's own condition. This results in higher stability and recognition accuracy for image processing and feature extraction in the high-temperature, high-heat-radiation environment of continuous casting. Finally, the system utilizes binocular vision data acquisition, combined with three-dimensional coordinate plane fitting of the four corner points and Bézier curve fitting techniques, to convert the detection results into quantifiable and traceable shape measurement results. This provides a reliable basis for adjusting continuous casting process parameters and quality control, thereby meeting the industrial site's demand for precise slab quality assessment.

[0189] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting the end face shape of a continuously cast slab based on binocular vision, characterized in that, Includes the following steps: Construct a binocular vision system to synchronously acquire images of the slab end faces on the continuous casting roller conveyor; Using a trained target detection model, the acquired images are identified to distinguish between the casting head area and the slab end face area to be tested, and the four corner points of the slab end face are located. When the end face region of the slab is identified and the corner points are complete, the region of interest of the end face is cropped, and the pixel-level contour of the end face of the slab is extracted from the region of interest using a trained contour segmentation model. Based on the calibration parameters of the binocular vision system and the principle of triangulation, the coordinates of the four corner points in three-dimensional space are calculated. Based on the three-dimensional coordinates of the four corner points, the spatial plane where the end face of the slab is located is fitted, and the homography mapping relationship between the spatial plane and the image pixel plane is established. Based on homography mapping, the pixel-level contours extracted from the image are mapped to the spatial plane to obtain end face contours with real physical dimensions. The mapped end face contour is segmented and smoothed using a curve fitting method, and then the width, height and edge shape deviation parameters of the slab end face are calculated.

2. The method according to claim 1, characterized in that, Constructing a binocular vision system to synchronously acquire images of the slab end faces on the continuous casting roller conveyor, specifically including: Two industrial cameras are fixedly installed in a sealed housing above the roller conveyor to form a binocular vision system. The cooling system continuously delivers cooling airflow into the camera's protective housing to maintain the camera's operating temperature in high-temperature environments. Image acquisition is automatically triggered based on changes in image grayscale. The acquired images are binarized, and the area of ​​the maximum connected region is calculated. When the area of ​​the maximum connected region exceeds a preset threshold and the area change between adjacent frames exceeds a set ratio, the slab is determined to have entered the field of view, and the binocular camera is triggered to acquire images synchronously.

3. The method according to claim 1, characterized in that, Before using a trained object detection model to recognize the acquired images, a vision system calibration step is also included: A calibration board with a concentric circle array is used; Acquire multiple images of the calibration plate from the left and right cameras in different poses; By constructing a frontal projection model of concentric circles, the coordinates of the extracted center feature points are optimized; Based on the optimized feature point coordinates, an iterative optimization strategy is adopted to jointly solve the internal parameters, distortion parameters, and relative pose relationship between the left and right cameras.

4. The method according to claim 1, characterized in that, The target detection model and contour segmentation model are trained based on the constructed slab end face dataset. The process of constructing the dataset includes: After automatically acquiring images containing slabs, manual screening and classification are performed, dividing them into casting head type images and end face type images; For images of casting heads, the casting head area is labeled; for images of end faces, the end face area and the positions of its four corner points are labeled, forming the training dataset for the target detection model. Select a subset of samples from end-face type images and perform pixel-level fine annotation on the complete contour of the slab end face to form a training dataset for the contour segmentation model.

5. The method according to claim 1, characterized in that, The trained object detection model is used to identify images, specifically as follows: The YOLO series target detection network is used to identify the end face area and the casting head area, and the bounding box of the end face area and the coordinates of the four corner points are output. The subsequent contour extraction and measurement process is executed only when the detection target is the end face of the slab and all four corner points are successfully located.

6. The method according to claim 1, characterized in that, The pixel-level contour of the slab end face is extracted using a trained contour segmentation model, specifically: An image segmentation network with a U-Net structure is used to process the region of interest on the cropped end face and output a binarized contour mask of the end face. Based on the corner point positions provided by the target detection model, the nearest contour point is searched on the extracted contour to accurately locate the actual corner point pixel coordinates.

7. The method according to claim 1, characterized in that, Based on the calibration parameters of the binocular vision system and the principle of triangulation, the coordinates of the four corner points in three-dimensional space are calculated as follows: Use the left camera coordinate system as the world coordinate system; For the matching corner pixel coordinates in the left and right images, a system of equations is established by combining the intrinsic parameters and distortion parameters of the left and right cameras, as well as the rotation matrix and translation vector of the right camera relative to the left camera. The least squares method is used to solve the system of equations to obtain the three-dimensional spatial coordinates of each corner point in the world coordinate system.

8. The method according to claim 1, characterized in that, Based on the three-dimensional coordinates of the four corner points, the spatial plane containing the end face of the slab is fitted, and a homography mapping relationship is established, specifically as follows: Calculate the centroids of the four corner points in 3D coordinates, and calculate the offset matrix of each corner point relative to the centroids; Perform singular value decomposition on the offset matrix, and take the left singular vector corresponding to the smallest singular value as the normal vector of the fitting plane; Each corner point is projected onto the fitting plane to obtain the projected points on the plane; Establish a two-dimensional coordinate system in the fitting plane with the centroid as the origin; The two-dimensional coordinates of the projection point in the two-dimensional coordinate system are calculated. By combining these two-dimensional coordinates with the corresponding pixel coordinates in the original image, the homography transformation matrix is ​​calculated.

9. The method according to claim 1, characterized in that, The mapped end face contour is segmented and smoothed using curve fitting, specifically as follows: The end face profile mapped onto the actual size plane is segmented along the edge direction at fixed intervals. For each segment of the contour point set, a third-order Bézier curve is used for fitting; Based on the fitted curve, the width and height of the slab end face are calculated, and the edge shape deviation is obtained by calculating the distance from the point on the contour to its corresponding edge reference line.

10. A continuous casting slab end face shape detection system based on binocular vision, characterized in that, The method for implementing any one of claims 1-9 includes a hardware module and a software module; The hardware module includes: The camera system module is used to synchronously acquire images of the end face of the slab running on the continuous casting roller conveyor; A cooling system module, connected to the camera system module, is used to deliver cooling airflow into the camera system module to ensure stable operation in high-temperature industrial environments. The software module is deployed on an industrial control computer, and the software module includes: The vision system calibration module is used to calibrate the parameters of the camera system module and establish the mapping relationship between image coordinates and three-dimensional space. The slab end face contour detection module is used to receive the acquired images and, based on the mapping relationship established by the vision system calibration module and the trained deep learning model, identify the slab end face region from the images, locate its four corner points, and extract the pixel-level contour of the end face. The slab end face contour shape measurement module is used to receive corner point and contour information, and calculate the actual width, height and edge shape deviation parameters of the slab end face according to the calibration parameters through three-dimensional reconstruction, plane fitting, homography mapping and curve fitting algorithms. The human-computer interaction module is used to control the system's start and stop, display the process and results of image acquisition, contour extraction, and parameter calculation in real time, and manage and export measurement data.

Images