Image recognition and positioning method and system for automatic feeding of steel strips of varying thicknesses

Abstract

This application relates to the field of image recognition technology, and discloses an image recognition and positioning method and system for automatic feeding of steel strips of varying thicknesses. First, it rapidly locates the approximate positions of the left and right edges of the steel strip through region of interest positioning. Next, for these two image blocks containing edge information, a dedicated edge regression network is used. This network not only learns from complex image features and directly regresses the precise mathematical model parameters of the edges, but also introduces a confidence evaluation mechanism. That is, while outputting edge parameters, the network also outputs a confidence level for the result. Finally, robust decision-making is performed based on the confidence level. When both edge detection results are reliable, the deviation is calculated normally. However, when the confidence level of a single edge is low due to occlusion or poor imaging, the reliable edge information and historical width data from the other side can be intelligently used to estimate the centerline position, thereby avoiding system failure due to local interference.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, and more specifically, to an image recognition positioning method and system for automatic feeding of steel strips of varying thicknesses. Background Technology

[0002] In modern industrial production, especially in the processing of materials such as steel and non-ferrous metals, precise feeding of steel strip is a crucial step in ensuring product quality, improving production efficiency, and reducing production costs. Whether rolling, shearing, slitting, or coiling, the steel strip needs to be accurately centered or positioned before entering the processing equipment to avoid problems such as deviation, scratches, and strip breakage, thereby ensuring the smooth progress of subsequent processes and the accuracy of the final product. However, in actual production, the thickness, width, and surface condition (such as oxide scale, oil stains, scratches, and irregular edges) of the steel strip often vary significantly. This is especially true for mixed-line production of steel strips of varying thicknesses, where the edge characteristics are complex, posing a significant challenge to traditional feeding and positioning methods.

[0003] Traditional steel strip feeding and positioning methods mainly rely on mechanical contact guiding devices, photoelectric sensors, or laser displacement sensors. Mechanical contact methods are prone to wear and slow response, and may damage the steel strip surface, making them unsuitable for steel strips of varying widths and thicknesses. While photoelectric or laser sensors are non-contact, their measurement accuracy is easily affected by surface reflection, color, dust, and changes in ambient lighting. Furthermore, they typically only detect localized areas, making it difficult to obtain overall edge information of the steel strip, especially when the strip edges have burrs, curls, or damage, significantly reducing the robustness and accuracy of the measurement results. In recent years, with the development of machine vision technology, image recognition-based steel strip positioning methods have been gradually introduced. These methods acquire images of the steel strip using industrial cameras, then utilize image processing algorithms to identify the strip edges and calculate its positional deviation. However, existing image recognition positioning schemes still face many challenges when handling steel strips of varying thicknesses. For example, traditional edge detection operators (such as Sobel and Canny) are sensitive to image noise and illumination changes, making it difficult to reliably extract complex or blurred steel strip edges. Methods based on template matching or feature point recognition lack adaptability to changes in steel strip edge morphology; their recognition accuracy drops sharply once the steel strip thickness or edge shape changes significantly. Furthermore, most existing solutions lack a mechanism to evaluate the reliability of recognition results during edge recognition. When some edges are occluded, illumination is uneven, or image quality is poor, the system may output incorrect positioning information, leading to feeding deviations or even production accidents. Simultaneously, processing the entire image imposes a significant computational burden, impacting real-time performance. Summary of the Invention

[0004] To address the aforementioned technical problems, this application is proposed. Embodiments of this application present an image recognition and positioning method and system for automatic feeding of steel strips of varying thicknesses, aiming to solve the problems of low positioning accuracy, poor adaptability, susceptibility to interference, and poor robustness in existing technologies when dealing with steel strips of different thicknesses and surface conditions.

[0005] According to one aspect of this application, an image recognition and localization method for automatic feeding of steel strips of varying thicknesses is provided, comprising: acquiring original image frames of the steel strip captured by an industrial camera; performing rapid region of interest (ROI) localization on the original image frames of the steel strip to obtain a left image block tensor and a right image block tensor; inputting the left image block tensor and the right image block tensor into a trained edge regression network to obtain left edge parameters, right edge parameters, left edge confidence, and right edge confidence; sampling the left and right ROI rectangles and determining the parameters based on the left edge parameters, right edge parameters, and homography. The matrix maps the sampling points from the local coordinate system to the world coordinate system to obtain the point sets of the left edge and the right edge in the world coordinate system. Based on the point sets of the left and right edges in the world coordinate system, the physical deviation of the steel strip centerline relative to the equipment baseline and the physical width of the steel strip centerline are calculated. Based on the confidence levels of the left and right edges, a robust decision is made on the physical deviation of the steel strip centerline relative to the equipment baseline to obtain the final deviation and the overall confidence level. The final deviation and the overall confidence level are input into the PLC controller to obtain the control signal.

[0006] According to another aspect of this application, an image recognition and positioning system for automatic feeding of steel strips of varying thicknesses is provided, comprising: a steel strip original image acquisition module for acquiring original image frames of the steel strip captured by an industrial camera; a steel strip image region of interest (ROI) localization module for rapidly locating the ROI of the original image frames of the steel strip to obtain a left image block tensor and a right image block tensor; an edge feature regression module for inputting the left image block tensor and the right image block tensor into a trained edge regression network to obtain left edge parameters, right edge parameters, left edge confidence, and right edge confidence; and an edge point set coordinate transformation module for sampling the left and right ROI rectangles and transforming the left and right ROI based on the left edge parameters and the right edge parameters. Edge parameters and homography matrices map sampling points from the local coordinate system to the world coordinate system to obtain the point sets of the left and right edges in the world coordinate system; the strip geometry parameter calculation module is used to calculate the physical deviation of the strip centerline relative to the equipment baseline and the physical width of the strip centerline based on the point sets of the left and right edges in the world coordinate system; the edge confidence decision module is used to make robust decisions on the physical deviation of the strip centerline relative to the equipment baseline based on the confidence levels of the left and right edges to obtain the final deviation and comprehensive confidence level; the deviation control signal generation module is used to input the final deviation and comprehensive confidence level into the PLC controller to obtain the control signal.

[0007] Compared with existing technologies, the image recognition and positioning method and system for automatic feeding of steel strips of varying thicknesses provided in this application firstly, rapidly locates the approximate positions of the left and right edges of the steel strip through region of interest positioning; then, for these two image blocks containing edge information, a dedicated edge regression network is used. This network not only learns from complex image features and directly regresses the precise mathematical model parameters of the edges, but also introduces a confidence evaluation mechanism, that is, the network outputs a confidence level for the result along with the edge parameters; finally, robust decision-making is performed based on the confidence level. When both edge detection results are reliable, the deviation is calculated normally. However, when the confidence level of a single edge is low due to occlusion or poor imaging, the reliable edge information and historical width data of the other side can be intelligently used to estimate the centerline position, thereby avoiding the failure of the entire system due to local interference. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 The illustration shows a schematic flowchart of an image recognition and positioning method for automatic feeding of steel strips of varying thicknesses according to an embodiment of this application.

[0010] Figure 2 The figure shows a schematic flowchart of step S2 in the image recognition and positioning method for automatic feeding of steel strips of different thicknesses according to an embodiment of this application.

[0011] Figure 3 The figure shows a schematic flowchart of step S3 in the image recognition and positioning method for automatic feeding of steel strips of different thicknesses according to an embodiment of this application.

[0012] Figure 4 The illustration shows a schematic flowchart of step S5 in the image recognition and positioning method for automatic feeding of steel strips of different thicknesses according to an embodiment of this application.

[0013] Figure 5 The figure shows a schematic flowchart of step S6 in the image recognition and positioning method for automatic feeding of steel strips of different thicknesses according to an embodiment of this application.

[0014] Figure 6 The figure shows a schematic block diagram of an image recognition and positioning system for automatic feeding of steel strips of varying thicknesses according to an embodiment of this application. Detailed Implementation

[0015] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0016] Figure 1 The illustration shows a schematic flowchart of an image recognition and positioning method and system for automatic feeding of steel strips of varying thicknesses according to an embodiment of this application. Figure 1As shown, this application provides an image recognition and localization method for automatic feeding of steel strips of varying thicknesses, comprising: S1, acquiring original image frames of the steel strip captured by an industrial camera; S2, rapidly locating regions of interest (ROIs) in the original image frames of the steel strip to obtain left and right image block tensors; S3, inputting the left and right image block tensors into a trained edge regression network to obtain left edge parameters, right edge parameters, left edge confidence, and right edge confidence; S4, sampling the left and right ROI rectangles and, based on the left edge parameters, right edge parameters, and homography moments... The sampling points are mapped from the local coordinate system to the world coordinate system to obtain the point set of the left edge and the point set of the right edge in the world coordinate system; S5, based on the point set of the left edge and the point set of the right edge in the world coordinate system, the physical deviation of the steel strip centerline relative to the equipment baseline and the physical width of the steel strip centerline are calculated; S6, based on the confidence level of the left edge and the confidence level of the right edge, a robust decision is made on the physical deviation of the steel strip centerline relative to the equipment baseline to obtain the final deviation and the comprehensive confidence level; S7, the final deviation and the comprehensive confidence level are input into the PLC controller to obtain the control signal.

[0017] For example, in step S1, the original image frame of the steel strip is acquired by an industrial camera. It should be understood that in modern industrial production, especially in the processing of materials such as steel and non-ferrous metals, precise alignment of the steel strip is crucial to ensuring product quality, improving production efficiency, and reducing losses. Traditional mechanical contact or simple photoelectric sensor methods often suffer from limitations such as wear, slow response, susceptibility to environmental interference, and inability to acquire comprehensive information. Image recognition-based solutions, however, can overcome these drawbacks, providing more flexible, accurate, and comprehensive data. Therefore, it is necessary to first acquire the original image frame of the steel strip using an industrial camera.

[0018] In one specific embodiment, acquiring raw image frames of the steel strip captured by an industrial camera includes: installing an industrial camera above the steel strip feeding line. The camera's mounting position and angle are precisely calibrated to ensure that its optical axis is approximately perpendicular to the steel strip surface, and that its field of view fully covers the steel strip at its widest possible width and its edges on both sides, reserving sufficient image information for subsequent edge recognition. Subsequently, the camera is connected to an industrial computer via a GigE Vision interface. GigE Vision is a gigabit Ethernet-based industrial camera interface standard that allows cameras to transmit image data at high speed, over long distances, and reliably via standard Ethernet hardware, greatly simplifying system wiring and integration. As the steel strip moves continuously on the production line and passes through the feeding area below the camera, the industrial computer acquires raw image frames of the steel strip in real time using image acquisition software (such as Halcon or OpenCV). The image acquisition software is responsible for communicating with the industrial camera, controlling the camera's imaging parameters such as exposure time, gain, and white balance, and capturing independent digital images from the continuous data stream sent by the camera according to the preset frame rate (i.e., the number of images captured per second) or external trigger signals. These image frames are then loaded into the memory of the industrial computer in real time for immediate use by subsequent image processing and analysis modules.

[0019] For example, in step S2, the region of interest (ROI) is quickly located on the original image frame of the steel strip to obtain the tensors of the left and right image blocks. It should be understood that, faced with high-resolution original images captured by industrial cameras, in order to balance real-time processing and the accuracy of final edge recognition, an efficient strategy must be adopted to concentrate subsequent complex computational resources on the most information-rich regions of the image. Running a sophisticated edge recognition algorithm directly on the entire high-resolution image would undoubtedly impose a huge computational burden, resulting in excessively slow processing speeds that would fail to meet the stringent real-time requirements of modern industrial production lines. Therefore, by quickly and accurately locking and cropping out the key regions containing the left and right edges of the steel strip—the so-called ROI—the computational load of subsequent fine processing can be significantly reduced, interference from irrelevant background regions can be eliminated, thereby significantly improving the overall system efficiency and recognition accuracy.

[0020] In one embodiment, such as Figure 2As shown, the method for rapidly locating the region of interest (ROI) of the original image frame of the steel strip to obtain the left and right image block tensors includes: S21, rapidly locating the ROI of the original image frame of the steel strip to obtain the left and right ROI rectangles; S22, extracting the left and right image blocks from the original image frame of the steel strip based on the left and right ROI rectangles; S23, performing size modulation based on bilinear interpolation on the left and right image blocks to obtain the size-normalized left and right image blocks; S24, normalizing the size-normalized left and right image blocks to obtain the left and right image block tensors.

[0021] Specifically, firstly, the regions of interest (ROIs) of the original image frame of the steel strip are quickly located to obtain the left and right ROI rectangles. To achieve rapid localization, this application employs a pre-trained lightweight object detection model, specifically the YOLOv5s model, renowned in the industry for its excellent balance between speed and accuracy. This model is not designed for precise pixel-level edge identification, but is specifically optimized for rapidly identifying the approximate positions of the left and right edges of the steel strip with extremely high efficiency after receiving the complete original image frame. Its lightweight nature is reflected in its streamlined architecture, which mainly consists of three parts: the backbone, the neck, and the head. The backbone uses an optimized CSPNet structure to efficiently extract multi-level features from the input image; the neck uses a PANet structure to cleverly fuse feature maps from different levels of the backbone, combining deep semantic information with shallow, precise localization information; finally, the detection head directly predicts on these fused feature maps, outputting a series of bounding box coordinates, confidence scores, and class probabilities.

[0022] To ensure the YOLOv5s model's ability to accurately identify steel strip edges, its pre-training process is crucial and extremely rigorous. First, a large-scale and diverse dataset of steel strip images needs to be constructed. This dataset must fully cover various situations that might be encountered in actual production, including steel strips of different thicknesses, widths, and materials, as well as complex conditions such as uneven lighting, surface oil stains, scratches, oxide scale, and partial occlusion. Then, each image in the dataset undergoes precise manual annotation. Using specialized tools, tight rectangular bounding boxes are drawn for the left and right edges of the steel strip, and category labels are assigned. The training process employs a transfer learning strategy to improve efficiency and performance. Specifically, based on the YOLOv5s model weights pre-trained on a large general dataset, fine-tuning is performed using our own annotated steel strip dataset. During fine-tuning, the model uses a composite loss function consisting of localization loss, object recognition loss, and classification loss to measure the difference between its predictions and the true annotations. Backpropagation and an optimizer are used to iteratively update the network parameters until the model achieves optimal performance on an independent validation set.

[0023] Through this meticulous design and training, the model ultimately learned to accurately identify the features of the steel strip edges under various complex backgrounds and interferences. When it receives a complete original image frame as input, it can quickly determine the possible regions where the left and right edges of the steel strip may exist by utilizing the deep features learned from a large amount of labeled data. The model outputs the coordinates of two bounding boxes, which precisely define the regions containing the left and right edges in the form of rectangles, referred to as the left ROI rectangle and the right ROI rectangle, respectively. Compared with traditional full-image scanning or coarse localization methods based on edge operators, this deep learning-based object detection method not only has an overwhelming advantage in speed but also achieves a qualitative leap in robustness. It can effectively cope with complex situations such as changes in lighting and differences in the surface texture of the steel strip, thereby ensuring the accuracy and real-time performance of ROI localization and providing high-quality, small-range input data for subsequent fine edge regression analysis.

[0024] After determining the ROI rectangles, the next step is to extract the left and right image blocks from the original image frame of the steel strip based on the left and right ROI rectangles. This operation is a direct image cropping process. Based on the coordinate information of the left and right ROI rectangles obtained in the previous step (i.e., the top-left pixel coordinates of each rectangle, as well as its width and height), the corresponding pixel data is precisely cropped from the original image frame. In this way, background information unrelated to the steel strip edges is effectively removed from the original image, retaining only the local image regions containing key edge features, thereby significantly reducing the amount of data for subsequent processing and avoiding unnecessary computational overhead.

[0025] Subsequently, to meet the uniform input size requirements of subsequent neural network models, the left and right image blocks are sized using bilinear interpolation to obtain size-standardized left and right image blocks. Since the size of the image blocks cropped by the object detection model may vary slightly due to minor differences in the position, width, or detection results of the steel strip, and the input layers of deep learning neural networks typically require data of a fixed size, size standardization is necessary. This application employs a bilinear interpolation algorithm for image scaling. Bilinear interpolation is a widely used smoothing scaling technique in image processing. It determines the pixel value of a new location pixel by calculating the weighted average of its four nearest neighbor pixels in the original image. Compared to simple nearest neighbor interpolation, this method generates smoother, visually better scaled images, effectively avoiding jagged artifacts or pixel distortion that may occur when adjusting image size, thus better preserving the detail information of the steel strip edges in the image block and ensuring the quality of the data input to the neural network.

[0026] Finally, to make the image data more suitable for neural network computation, the size-normalized left and right image patches are normalized to obtain the left and right image patch tensors. Here, normalization refers to converting the color value of each pixel in the image patch from the usual integer range of [0, 255] to the floating-point range of [0, 1] through a linear mapping (e.g., simply dividing by 255.0). This normalization process helps accelerate the convergence speed of neural network model training and improves model stability because it prevents gradient explosion or vanishing problems caused by excessively large or small input data.

[0027] For example, in step S3, the tensors of the left and right image blocks are input into the trained edge regression network to obtain the left edge parameters, right edge parameters, left edge confidence, and right edge confidence. It should be understood that traditional steel strip edge detection methods often rely on fixed image processing algorithms, such as edge detection based on operators like Sobel and Canny, or techniques based on threshold segmentation and template matching. These methods may perform reasonably well under ideal lighting and surface conditions, but they are vulnerable to complex conditions common in industrial production. For example, steel strips of varying thicknesses, edge morphologies, and surface conditions, as well as interference factors such as uneven lighting, surface oil, scratches, oxide scale, and local occlusion, exhibit weak anti-interference capabilities and poor robustness, easily leading to false detections, missed detections, or unstable detection results. More importantly, these traditional methods typically only provide pixel-level edge positions, lacking a quantitative assessment of the reliability of the detection results. That is, they cannot determine how reliable the currently detected edge is, making it difficult for subsequent control systems to make intelligent decisions under abnormal conditions.

[0028] In one embodiment, such as Figure 3 As shown, the process of inputting the tensors of the left and right image patches into a trained edge regression network to obtain left edge parameters, right edge parameters, left edge confidence, and right edge confidence includes: S31, inputting the tensors of the left and right image patches into a lightweight convolutional neural network model to obtain left image semantic feature encoding vectors and right image semantic feature encoding vectors; S32, inputting the left image semantic feature encoding vectors and right image semantic feature encoding vectors into a regression head to obtain left edge parameters and right edge parameters; S33, inputting the left image semantic feature encoding vectors and right image semantic feature encoding vectors into a confidence head to obtain left edge confidence and right edge confidence.

[0029] Specifically, firstly, the left and right image patch tensors are input into a lightweight convolutional neural network model to obtain the semantic feature encoding vectors for the left and right images. In this sub-step, the left and right image patch tensors (i.e., image data that has been size-normalized and standardized) generated in the previous preprocessing stage are fed in parallel into a lightweight convolutional neural network model that serves as the backbone for feature extraction. A lightweight model (e.g., the MobileNetV3-Small architecture) is chosen to minimize computational complexity and memory usage while maintaining recognition accuracy, thus meeting the real-time requirements of industrial environments. This CNN model, through its combination of multiple convolutional layers, pooling layers, and activation functions, can extract abstract features from the original pixels layer by layer, ranging from low-level (e.g., edge texture, gradients, corner points) to high-level (e.g., edge continuity, overall shape, and distinction from the background). As data is passed layer by layer through the network, these features become increasingly abstract and semantic. Ultimately, the model transforms each input image patch into a fixed-length numerical vector containing rich semantic information—the image semantic feature encoding vector. This vector is a highly condensed and essential representation of the original image patch content, capturing the deep patterns in the image most relevant to steel strip edge recognition.

[0030] After obtaining the semantic feature encoding vectors, the network performs two different tasks in parallel, each completed by a separate head. The first task is to input the semantic feature encoding vectors from the left and right images into the regression head to obtain the left and right edge parameters. This "regression head" is part of the network architecture and typically consists of one or more fully connected layers. It receives the semantic feature encoding vectors output by the feature extractor as input and maps these abstract features to concrete edge parameters using the weights and biases learned during training. For example, if the edge is modeled as a straight line, the regression head will output three floating-point numbers. Essentially, this process associates abstract image features with precise mathematical models describing edges, achieving a direct conversion from image to geometric parameters.

[0031] Simultaneously, a second task is performed: inputting the semantic feature encoding vectors of the left and right images into the confidence head to obtain the confidence scores for the left and right edges. Similar to the regression head, the confidence head is also a separate network branch, typically composed of fully connected layers, with a possible sigmoid activation function at its end to ensure the output value is between 0 and 1. It also receives semantic feature encoding vectors as input, but its goal is to output a single confidence score. This branch is specifically trained to identify the presence of clear, complete, and reliable edge features in the image, and accordingly assigns a score between 0 and 1. For example, when the edges in an image patch are clearly visible, unobstructed, and free of significant noise, the confidence head outputs a value close to 1; while when the edges are blurred, obstructed, or subject to severe interference, it outputs a value close to 0.

[0032] The training process of the edge regression network is crucial to its high accuracy and robustness. Training such a deep learning model requires a large and high-quality dataset. This dataset must contain numerous images of steel strips under various working conditions, and each image needs precise manual annotation. For edge parameter annotation, professionals need to accurately depict the left and right edges of the steel strip on the image and convert them into corresponding mathematical parameters (e.g., fitting a straight line using the least squares method). For edge confidence annotation, a reliability score needs to be assigned to each edge based on image quality and edge sharpness, or it needs to be classified as "reliable" or "unreliable." The diversity of the dataset is essential; it must cover steel strips of different thicknesses, different surface conditions (e.g., glossy, oxidized, oily, scratched), different lighting conditions (e.g., uniform, localized shadows, reflections), and various possible interferences (e.g., partial occlusion, background clutter) to ensure the model has strong generalization ability.

[0033] During training, the model is optimized through multi-task learning. This means that the regression head and confidence head are trained simultaneously, sharing the same feature extractor. The total loss function is a weighted sum of the regression loss and the confidence loss. The regression loss (e.g., mean squared error (MSE) or L1 loss) measures the difference between the network's predicted marginal parameters and the true labeled parameters, aiming to minimize this difference. The confidence loss (e.g., binary cross-entropy loss or mean squared error) measures the difference between the network's predicted confidence and the true confidence, aiming to enable the network to accurately assess the reliability of its own predictions. Through backpropagation algorithms and optimizers (such as Adam or SGD), the network continuously adjusts its internal weights and biases based on the gradient of the loss function until the model achieves satisfactory performance on the training set and demonstrates good generalization ability on independent validation sets, avoiding overfitting. To accelerate training and improve performance, transfer learning is often employed, which uses a lightweight CNN model pre-trained on a large-scale image dataset (such as ImageNet) as the initial weights for the feature extractor, and then fine-tunes it to adapt it to the specific task of steel strip edge detection.

[0034] Here, for the edge parameters and confidence scores regressed by the edge regression network, since the robustness of subsequent steps, namely coordinate transformation and control decision-making, needs to be considered, the model is expected to express the estimation uncertainty. For example, suppose the edge of the steel strip is very blurry or highly reflective in a small area. In reality, there may be multiple curves and their confidence scores that can fit this blurry edge. This uncertainty, which originates from the noise or ambiguity of the data itself, is usually called random uncertainty. When this parameter uncertainty is used for coordinate transformation and control decision-making, this uncertainty will continue to propagate, thus affecting the accuracy of the parameters and the control decision. Therefore, a negative log-likelihood loss function for the multivariate Gaussian covariance distribution is further introduced during the training process of the edge regression network.

[0035] In another embodiment, the training process of the edge regression network includes: determining the regression loss value for the regression head; determining the confidence loss value for the confidence head; performing multivariate Gaussian covariance transformation on the left and right image semantic feature encoding vectors to obtain Gaussian distributed left and right image semantic feature encoding vectors; calculating the Gaussian covariance matrix based on the Gaussian distributed left and right image semantic feature encoding vectors; calculating the negative log-likelihood loss function value for the multivariate Gaussian covariance distribution based on the Gaussian covariance matrix, the Gaussian distributed left and right image semantic feature encoding vectors; and training the edge regression network by a weighted sum of the regression loss, confidence loss, and negative log-likelihood loss function value.

[0036] Specifically, the training process first focuses on the model's fundamental predictive ability, namely determining the regression loss value for the regression head and the confidence loss value for the confidence head. Specifically, a function such as smoothed L1 loss is used to measure the difference between the model's predicted marginal parameters and the manually labeled true parameters, driving the network to continuously reduce this difference through backpropagation. This ensures the accuracy of the model's predictions, enabling the network to output marginal equations that closely match the actual positions of the steel strip under ideal conditions.

[0037] Next, it is expected that the semantic feature encoding vectors of the left and right images output by the lightweight convolutional neural network model will have a joint probability distribution descriptiveness. Therefore, for the semantic feature encoding vector of the left image, for example, denoted as... and the semantic feature encoding vector of the image on the right, for example, denoted as Multivariate Gaussian covariance transformation is performed to obtain Gaussian distributed semantic feature encoding vectors for the left and right sides of the image, i.e.: ;in, It represents each feature value of the semantic feature encoding vector of the left image and the semantic feature encoding vector of the right image, and and These are the mean and variance of all feature values ​​of the semantic feature encoding vectors of the left and right images, respectively. Let each eigenvalue represent a Gaussian-distributed semantic feature encoding vector for the left and right images, then the Gaussian covariance matrix is: ;in, This represents vector subtraction. Represents matrix multiplication. Represents the transpose of a vector. This represents the Gaussian-distributed semantic feature encoding vector of the left side of the image. This represents the Gaussian-distributed semantic feature encoding vector of the right-hand image. It is the positional mean feature vector of the Gaussian-distributed semantic feature encoding vector of the left image and the Gaussian-distributed semantic feature encoding vector of the right image, and both vectors are in column vector form. Let represent the Gaussian covariance matrix. Therefore, the Gaussian covariance matrix represents the joint distribution of image semantic coding features based on uncertainty and its correlation.

[0038] Then, based on the Gaussian covariance matrix, the Gaussian-distributed left-side image semantic feature encoding vector, and the Gaussian-distributed right-side image semantic feature encoding vector, the negative log-likelihood loss function value for the multivariate Gaussian covariance distribution is calculated, i.e.: ;in, Denotes the inverse of the Gaussian covariance matrix. and Let L2 and LF be the matrix norms, respectively. For bias hyperparameters, This represents the negative log-likelihood loss function for a multivariate Gaussian covariance distribution. This loss function, through its first term, encourages the semantic feature encoding vectors of the left and right images to approximate the consistency of their Gaussian covariance, while simultaneously demonstrating that uncertainty increases or decreases in the same direction. The second term simulates the uncertainty volume, minimizing its uncertainty information gain. For example, when edges are blurred, the multivariate Gaussian covariance distribution... The norm representation would be too large, thus uncertainty quantization is used to control the encoding of image semantic features. Finally, the regression loss, confidence loss, and negative log-likelihood loss function values ​​are weighted and summed, and this total loss value is used to update the gradient and optimize the parameters of the entire edge regression network.

[0039] For example, in step S4, the left and right ROI rectangles are sampled, and the sampled points are mapped from the local coordinate system to the world coordinate system based on the left edge parameters, right edge parameters, and homography matrix to obtain the point set of the left edge and the point set of the right edge in the world coordinate system. It should be understood that in industrial automation, especially in steel strip feeding systems, actuators such as PLC controllers require the actual physical deviation of the steel strip relative to the equipment baseline, rather than the pixel deviation in the image. Because images captured by industrial cameras have perspective distortion, the pixel distance in the image cannot be directly equated to the actual physical distance, and there is a complex geometric transformation relationship between the image coordinate system and the actual physical coordinate system. Therefore, achieving accurate and high-precision conversion from the image coordinate system to the physical world coordinate system is a key step in ensuring that the entire automatic feeding system can effectively correct deviations and guarantee product quality. This conversion process not only eliminates image distortion but also accurately projects the two-dimensional image information into a three-dimensional (or simplified to a two-dimensional plane) physical space, thereby providing reliable data with practical physical meaning for subsequent physical deviation and width calculations.

[0040] In one embodiment, sampling is performed on the left and right ROI rectangles, and the sampling points are mapped from the local coordinate system to the world coordinate system based on the left edge parameters, the right edge parameters, and the homography matrix to obtain the point set of the left edge in the world coordinate system and the point set of the right edge in the world coordinate system. This includes: mapping the sampling points in the left and right ROI rectangles from the local coordinate system to the image pixel coordinate system based on the left edge parameters and the right edge parameters to obtain the point set of the left edge in the image coordinate system and the point set of the right edge in the image coordinate system; and mapping the point set of the left edge in the image coordinate system and the point set of the right edge in the image coordinate system from the image pixel coordinate system to the world coordinate system based on the homography matrix to obtain the point set of the left edge in the world coordinate system and the point set of the right edge in the world coordinate system.

[0041] Specifically, firstly, based on the left and right edge parameters, the sampling points in the left and right ROI rectangles are mapped from the local coordinate system to the image pixel coordinate system to obtain the point sets of the left edge and the right edge in the image coordinate system. In this sub-step, the left and right edge parameters output by the edge regression network in the previous stage are used. Within the respective ROI rectangle, uniform sampling is performed along the edge lines defined by these parameters. Specifically, within the local coordinate system of the ROI, a certain number (e.g., 100) of sampling points are calculated based on the mathematical model of the edge (e.g., for a straight line, the X or Y coordinates can be fixed, and then the other coordinate can be calculated according to the equation). These sampling points are initially local coordinates relative to the top-left corner of their respective ROIs. Subsequently, in order to convert these local coordinates into absolute positions in the entire original image frame, these local coordinates are added to the starting coordinates of the corresponding ROI rectangle in the original image pixel coordinate system (i.e., the top-left corner pixel coordinates of the ROI). For example, if the coordinates of the top-left corner of an ROI in the original image are (roi... x ,roi y The local coordinates of a sampling point within the ROI are (local... x ,local y If ), then its coordinates in the image pixel coordinate system are (roi) x +local x ,roi y +local y By performing the above transformation on all sampling points, we obtain the set of discrete points for the left and right edges in the entire image pixel coordinate system. This step ensures that subsequent physical coordinate mapping can be based on the absolute position in the image.

[0042] Secondly, a process based on the homography matrix is ​​performed to map the point sets of the left and right edges in image coordinates from the image pixel coordinate system to the world coordinate system, thus obtaining the point sets of the left and right edges in world coordinates. In this sub-step, a 3x3 homography matrix H, pre-obtained through camera calibration, is used. This homography matrix is ​​calculated during system installation and debugging by photographing a calibration board with known physical dimensions and positions, and using computer vision algorithms (such as the Zhang Zhengyou calibration method). The calibration process establishes a precise geometric correspondence between the camera image plane and the physical plane where the steel strip is located. It can correct the camera's internal parameters (such as focal length, principal point, and distortion coefficient) and external parameters (the camera's position and orientation relative to the world coordinate system), thereby achieving accurate projection from a two-dimensional image to three-dimensional physical space. Specifically, each sampling point of the left and right edges in the image pixel coordinate system obtained in the previous step is mapped to its corresponding point in the physical world coordinate system through matrix multiplication with the homography matrix `H`.

[0043] For example, in step S5, based on the point sets of the left and right edges in world coordinates, the physical deviation of the steel strip centerline relative to the equipment baseline and the physical width of the steel strip centerline are calculated. It should be understood that in industrial production, simply knowing the physical coordinates of the steel strip edges is insufficient; more importantly, it is necessary to understand whether the overall width of the steel strip conforms to specifications and whether its centerline is precisely aligned with the preset baseline of the production equipment. Only by accurately calculating the actual width of the steel strip and the offset of its centerline relative to the preset baseline of the equipment can a precise correction control signal be generated for the subsequent PLC controller, thereby ensuring that the steel strip maintains correct alignment during feeding and avoiding production accidents, product quality problems, or equipment wear caused by deviation. This step serves as a bridge connecting visual perception and actual physical control, transforming abstract coordinate data into physical indicators that can be directly used for industrial decision-making.

[0044] In one embodiment, such as Figure 4 As shown, based on the point sets of the left edge and the right edge in world coordinates, the physical deviation of the steel strip centerline relative to the equipment baseline and the physical width of the steel strip centerline are calculated, including: S51, extracting a set of paired {left edge points, right edge points} from the point sets of the left edge and the right edge in world coordinates; S52, calculating a set of widths and a set of deviations based on the set of paired {left edge points, right edge points}; S53, calculating the mean of the set of widths as the physical width of the steel strip centerline; S54, calculating the mean of the set of widths as the physical deviation of the steel strip centerline relative to the equipment baseline.

[0045] Specifically, first, pairs of {left edge points, right edge points} are extracted from the point sets of the left and right edges in world coordinates. In this sub-step, the point sets of the left and right edges obtained in the previous stage are matched according to their Y-axis coordinates (i.e., coordinates along the length of the steel strip). Since the edges of the steel strip are usually continuous and straight, corresponding left and right edge points can be found at the same or similar Y-axis positions. Pairing is achieved by traversing one point set and finding the point with the closest Y-coordinate in the other set. Typically, since the sampling points are uniformly distributed, pairing can be done directly by index, forming a set of multiple (e.g., 100 pairs) of {left edge points, right edge points} with the same or similar Y-axis coordinates.

[0046] Next, the system will perform calculations on the set of paired {left edge points, right edge points}, calculating both the width and the deviation. For each pair of paired points extracted in the previous step, the system performs two key geometric calculations. The first is to calculate the local width of that point. This represents the absolute distance between the left and right edges of the steel strip along the X-axis at a specific Y-coordinate along the length of the steel strip, i.e., the instantaneous width at that point. The second is to calculate the local centerline X-coordinate of that point. This represents the position of the geometric center point along the width direction of the steel strip on the X-axis of the world coordinate system at the same Y-coordinate. Based on this, the local deviation of the point is calculated.

[0047] Next, the mean of the set of widths is calculated as the physical width of the steel strip's centerline. To obtain the overall physical width of the steel strip, all local width values ​​in the set of widths are statistically averaged. By averaging these local widths, the effects of local measurement errors, minor irregularities at the steel strip edges (such as burrs or slight waviness), and camera noise can be effectively eliminated, resulting in a more stable, accurate, and representative physical width value for the steel strip. This average width value is the actual width of the steel strip at the current moment and can be used for quality control and production parameter adjustment.

[0048] Finally, the mean of the set of deviations is calculated as the physical deviation of the strip centerline relative to the equipment baseline. Similar to the width calculation, all local deviation values ​​in the set of deviations are statistically averaged. This average represents the overall physical offset of the strip centerline relative to the equipment baseline at the current moment. A positive value indicates that the strip centerline is biased to one side of the equipment baseline (e.g., the right side), and a negative value indicates biased to the other side (e.g., the left side). This average deviation value is the direct basis for subsequent correction operations by the PLC controller, providing precise instructions on which direction and how far the strip needs to move.

[0049] For example, in step S6, a robust decision is made on the physical deviation of the steel strip centerline relative to the equipment baseline based on the confidence levels of the left and right edges to obtain the final deviation and overall confidence level. It should be understood that in actual industrial production environments, the imaging conditions of the steel strip are often complex and variable, far from ideal. For example, the steel strip surface may have defects such as oxide scale, oil stains, scratches, and reflections, or during the feeding process, there may be local occlusion (such as operators passing by or equipment components obstructing the view), uneven lighting (such as local shadows or strong light reflections), etc. These interfering factors may cause the image features of one or both edges to be blurred or incomplete, thus making the reliability of the edge parameters output by the edge regression network in the previous stage, even though they have undergone preliminary calculation, potentially inconsistent. If the system blindly relies on these potentially erroneous measurements, it will directly affect the accuracy of subsequent control signals, and may even lead to steel strip deviation, equipment damage, or production interruption. Therefore, this step introduces a quantitative assessment of the reliability of edge detection results (i.e., confidence level) and intelligently adjusts the deviation calculation strategy based on the assessment results. This ensures that even if some data is unreliable, the system can still output a stable, accurate, and highly robust final deviation value, thereby significantly improving the reliability and adaptability of the entire automatic feeding system and avoiding system failures caused by environmental changes or occasional anomalies.

[0050] In one embodiment, such as Figure 5 As shown, based on the confidence levels of the left and right edges, a robust decision is made on the physical deviation of the steel strip centerline relative to the equipment baseline to obtain the final deviation and a comprehensive confidence level. This includes: S61, taking the minimum of the confidence levels of the left and right edges as the comprehensive confidence level; S62, if both the confidence levels of the left and right edges are greater than a preset threshold, determining the physical deviation of the steel strip centerline relative to the equipment baseline as the final deviation; S63, if the confidence level of the left edge is less than the preset threshold but the confidence level of the right edge is greater than the preset threshold, calculating the final deviation based on the point set of the right edge in world coordinates and the previously measured physical width of the steel strip centerline; S64, if the confidence level of the right edge is less than the preset threshold but the confidence level of the left edge is greater than the preset threshold, calculating the final deviation based on the point set of the left edge in world coordinates and the previously measured physical width of the steel strip centerline.

[0051] Specifically, the minimum of the confidence scores for the left and right edges is taken as the overall confidence score. This operation aims to evaluate the overall reliability of the measurement in the most conservative way. The logic is that the reliability of the entire measurement result will not be higher than its least reliable part. If the detection confidence score of one edge is very low, even if the confidence score of the other edge is very high, it means that there may be a large uncertainty in the calculation of the entire steel strip centerline. Therefore, the overall confidence score should reflect this lowest level of reliability.

[0052] Subsequently, based on the comparison between the confidence scores of the left and right edges and preset thresholds, different decision branches are initiated to address various actual working conditions. The first case is where, if both the confidence scores of the left and right edges are greater than the preset threshold (e.g., 0.8), the physical deviation of the steel strip centerline relative to the equipment baseline is determined as the final deviation. This is an ideal situation, indicating that the steel strip edges on both sides are clearly and reliably detected, and the edge regression network has high confidence in its prediction results. In this case, the physical deviation of the steel strip centerline relative to the equipment baseline calculated based on the left and right edge point sets in the previous step is considered highly accurate and can be directly output as the final deviation value without any correction or extrapolation. This ensures that the system can provide the most accurate positioning information under optimal conditions.

[0053] The second scenario involves calculating the final deviation based on the right edge's point set in world coordinates and the previously measured physical width of the steel strip's centerline, if the confidence level of the left edge is less than a preset threshold but the confidence level of the right edge is greater than a preset threshold. This scenario typically occurs when the left edge is partially obstructed, poorly lit, or has defects on the steel strip edge (such as severe damage or oxide scale peeling), resulting in low detection confidence, while the right edge detection result remains clear and reliable. To provide an accurate deviation value in this case, reliable information about the right edge and the known width of the steel strip are used for estimation. In one embodiment, if the left edge's confidence level is less than a preset threshold but the right edge's confidence level is greater than a preset threshold, calculating the final deviation based on the right edge's point set in world coordinates and the previously measured physical width of the steel strip's centerline includes: calculating the mean X-axis coordinate of the right edge's point set in world coordinates to obtain the right edge position. This is done by averaging the X-coordinates of all sampled points of the right edge in the world coordinate system to obtain a stable and representative average right edge position. The final deviation is obtained by subtracting half of the physical width of the previously measured steel strip centerline from the right edge position. Since the width of the steel strip is typically relatively stable over short periods (unless a roll change or special process occurs), even if the left edge is unreliable, the X-coordinate of the steel strip centerline can be estimated by subtracting half the steel strip width from the reliable right edge position, thus calculating its deviation relative to the equipment baseline. This method utilizes prior knowledge of the steel strip width, effectively compensating for the lack of information on one side.

[0054] The third scenario involves calculating the final deviation based on the set of points on the left edge in world coordinates and the previously measured physical width of the steel strip centerline if the confidence level of the right edge is less than a preset threshold but the confidence level of the left edge is greater than a preset threshold. In one embodiment, this calculation includes: calculating the average X-axis coordinate of the set of points on the left edge in world coordinates to obtain the left edge position, i.e., averaging the X-coordinates of all sampled points on the left edge in the world coordinate system to obtain a stable and representative average left edge position. The final deviation is obtained by subtracting half the previously measured physical width of the steel strip centerline from the left edge position. By adding half the steel strip width to the reliable left edge position, the X-coordinate of the steel strip centerline can be estimated, and its deviation relative to the equipment baseline can be calculated. These two unilateral extrapolation strategies significantly enhance the system's robustness in cases of missing or unreliable information.

[0055] For example, in step S7, the final deviation and overall confidence level are input into the PLC controller to obtain a control signal. It should be understood that in industrial automated production, the vision system is merely the eyes and brain, while the PLC controller is the arms and legs, responsible for implementing the brain's decisions. Only by effectively transmitting the calculated accurate deviation information and its reliability assessment to the PLC controller, and having it generate corresponding control signals according to the preset control strategy, can the automatic feeding, positioning, and correction of the steel strip be truly achieved, ensuring smooth production, stable product quality, and safe equipment operation.

[0056] In one embodiment, inputting the final deviation and overall confidence level into the PLC controller to obtain a control signal includes: multiplying the overall confidence level by the proportional coefficient of the PLC controller to obtain an updated proportional coefficient.

[0057] Specifically, firstly, the overall confidence level is multiplied by the proportional gain of the PLC controller to obtain an updated proportional gain. In this sub-step, the overall confidence level obtained from the robustness decision in the previous stage (a value between 0 and 1, reflecting the reliability of the measurement result) is used as a dynamic weighting factor and applied to the preset proportional gain within the PLC controller. The proportional gain is an important parameter in the PID control algorithm, determining the strength of the controller's response to deviations: the larger the proportional gain, the stronger the controller's response to the same deviation. By multiplying the overall confidence level by the initial proportional gain, a dynamically adjusted updated proportional gain can be obtained. This design is a significant advantage of this application. Its purpose is that when the overall confidence level of image recognition is high (e.g., close to 1), the updated proportional gain will be close to or equal to the initial proportional gain, making the controller's response to deviations more sensitive and direct, thereby achieving fast and accurate correction; while when the overall confidence level is low (e.g., under extremely poor imaging conditions, the confidence level may drop to 0.5 or even lower), the updated proportional gain will decrease accordingly, making the controller's response to deviations more conservative and slower. This adaptive adjustment mechanism can effectively avoid overcorrection, system oscillation, or unnecessary mechanical wear caused by unreliable measurement data, thereby significantly enhancing the robustness, stability, and safety of the entire control system.

[0058] Subsequently, control signals are generated. In this sub-step, the PLC controller receives two core inputs from the vision system: the final deviation after robust decision-making (e.g., +15 mm, indicating that the steel strip centerline deviates from the equipment baseline by 15 mm) and the updated proportional coefficient calculated in the previous step (e.g., 0.46). The PLC controller has advanced preset control algorithms, the most common and widely used being the PID (Proportional-Integral-Derivative) control algorithm. Using these input parameters, combined with its internal control logic and the current system state (e.g., the steel strip's running speed, historical deviation trends, etc.), the PLC calculates precise control signals for adjusting the steel strip feeding mechanism. These control signals can take various forms, such as analog voltage signals (for controlling the speed / position of frequency converters or servo drives), pulse width modulation (PWM) signals (for controlling motor speed), digital signals (for switching actuators), or digital commands sent via an industrial bus. These signals are sent in real-time to the corresponding actuators, such as laterally moving guide rollers, servo motor drives for uncoilers or coilers, hydraulic cylinders, etc. These drives precisely adjust the lateral position of the steel strip based on the received control signals, gradually bringing its centerline back to the equipment's baseline, thus achieving automatic centering of the steel strip. The entire process forms a high-speed, high-precision closed-loop feedback control system, ensuring the stability and accuracy of the steel strip throughout the feeding process.

[0059] In summary, the image recognition and localization method for automatic feeding of steel strips of varying thicknesses provided in this application firstly rapidly locates the approximate positions of the left and right edges of the steel strip through region of interest localization. Next, for these two image blocks containing edge information, a dedicated edge regression network is used. This network not only learns from complex image features and directly regresses the precise mathematical model parameters of the edges, but also introduces a confidence evaluation mechanism. That is, while outputting edge parameters, the network also outputs a confidence level for the result. Finally, robust decision-making is performed based on the confidence level. When both edge detection results are reliable, the deviation is calculated normally. However, when the confidence level of a single edge is low due to occlusion or poor imaging, the reliable edge information and historical width data from the other side can be intelligently used to estimate the centerline position, thereby avoiding system failure due to local interference.

[0060] This application also provides an image recognition and positioning system for automatic feeding of steel strips of varying thicknesses, used to execute the aforementioned image recognition and positioning method for automatic feeding of steel strips of varying thicknesses, such as... Figure 6 As shown, the image recognition and positioning system 600 for automatic feeding of steel strips of varying thicknesses includes: a steel strip original image acquisition module 610, used to acquire original image frames of the steel strip captured by an industrial camera; a steel strip image region of interest (ROI) localization module 620, used to quickly locate the ROI of the original steel strip image frame to obtain the left image block tensor and the right image block tensor; an edge feature regression module 630, used to input the left image block tensor and the right image block tensor into a trained edge regression network to obtain left edge parameters, right edge parameters, left edge confidence, and right edge confidence; and an edge point set coordinate transformation module 640, used to sample the left and right ROI rectangles and transform the coordinates based on the left edge parameters and right edge parameters. The homography matrix maps the sampling points from the local coordinate system to the world coordinate system to obtain the point set of the left edge and the point set of the right edge in the world coordinate system; the strip geometric parameter calculation module 650 is used to calculate the physical deviation of the strip centerline relative to the equipment baseline and the physical width of the strip centerline based on the point set of the left edge and the point set of the right edge in the world coordinate system; the edge confidence decision module 660 is used to make a robust decision on the physical deviation of the strip centerline relative to the equipment baseline based on the confidence of the left edge and the confidence of the right edge to obtain the final deviation and the comprehensive confidence; the deviation control signal generation module 670 is used to input the final deviation and the comprehensive confidence into the PLC controller to obtain the control signal.

[0061] This application also provides a computer-readable storage medium storing computer program code. When the computer program code is run on a computer, the computer executes the above-described related method steps to implement the image recognition and positioning method for automatic feeding of thick and thin steel strips provided in the above embodiments.

[0062] This application also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned steps to implement the image recognition and positioning method for automatic feeding of thick and thin steel strips provided in the above embodiments.

[0063] In this application, the system, computer-readable storage medium, or computer program product provided in the embodiments are all used to execute the corresponding methods provided above. Therefore, the beneficial effects that can be achieved can be referred to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

[0064] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments.

[0065] The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous. The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. An image recognition and positioning method for automatic feeding of steel strips of varying thicknesses, characterized in that, include: Acquire raw image frames of the steel strip captured by an industrial camera; The method for rapidly locating regions of interest (ROIs) in the original image frame of the steel strip to obtain left and right image block tensors includes: rapidly locating ROIs in the original image frame of the steel strip to obtain left and right ROI rectangles; extracting left and right image blocks from the original image frame of the steel strip based on the left and right ROI rectangles; performing size modulation on the left and right image blocks based on bilinear interpolation to obtain size-normalized left and right image blocks; and normalizing the size-normalized left and right image blocks to obtain left and right image block tensors. Input the tensors of the left and right image patches into the trained edge regression network to obtain the left edge parameters, right edge parameters, left edge confidence, and right edge confidence; The left and right ROI rectangles are sampled, and the sampled points are mapped from the local coordinate system to the world coordinate system based on the left edge parameters, the right edge parameters, and the homography matrix to obtain the point set of the left edge in the world coordinate system and the point set of the right edge in the world coordinate system. Based on the point set of the left edge and the point set of the right edge in world coordinates, calculate the physical deviation of the steel strip centerline relative to the equipment baseline and the physical width of the steel strip centerline. Based on the confidence levels of the left and right edges, a robust decision is made regarding the physical deviation of the steel strip centerline relative to the equipment baseline to obtain the final deviation and a comprehensive confidence level. This includes: taking the minimum of the confidence levels of the left and right edges as the comprehensive confidence level; if both the confidence levels of the left and right edges are greater than a preset threshold, determining the physical deviation of the steel strip centerline relative to the equipment baseline as the final deviation; if the confidence level of the left edge is less than the preset threshold but the confidence level of the right edge is greater than the preset threshold, calculating the final deviation based on the point set of the right edge in world coordinates and the previously measured physical width of the steel strip centerline; if the confidence level of the right edge is less than the preset threshold but the confidence level of the left edge is greater than the preset threshold, calculating the final deviation based on the point set of the left edge in world coordinates and the previously measured physical width of the steel strip centerline. The final deviation and overall confidence level are input into the PLC controller to obtain the control signal.

2. The image recognition and positioning method for automatic feeding of steel strips of varying thicknesses according to claim 1, characterized in that, The tensors of the left and right image patches are input into the trained edge regression network to obtain the left edge parameters, right edge parameters, left edge confidence, and right edge confidence, including: The tensors of the left and right image blocks are input into a lightweight convolutional neural network model to obtain the semantic feature encoding vectors of the left and right images. Input the semantic feature encoding vectors of the left and right images into the regression head to obtain the left edge parameters and the right edge parameters; Input the semantic feature encoding vectors of the left and right images into the confidence head to obtain the confidence scores of the left and right edges.

3. The image recognition and positioning method for automatic feeding of steel strips of varying thicknesses according to claim 2, characterized in that, The left and right ROI rectangles are sampled, and the sampled points are mapped from the local coordinate system to the world coordinate system based on the left edge parameters, right edge parameters, and homography matrix to obtain the point set of the left edge and the point set of the right edge in the world coordinate system, including: Based on the left and right edge parameters, the sampling points in the left and right ROI rectangles are mapped from the local coordinate system to the image pixel coordinate system to obtain the point set of the left edge and the point set of the right edge in the image coordinate system; and Based on the homography matrix, the point set of the left edge in image coordinates and the point set of the right edge in image coordinates are mapped from the image pixel coordinate system to the world coordinate system to obtain the point set of the left edge in world coordinates and the point set of the right edge in world coordinates.

4. The image recognition and positioning method for automatic feeding of steel strips of varying thicknesses according to claim 1, characterized in that, Based on the point sets of the left and right edges in world coordinates, the physical deviation of the steel strip centerline relative to the equipment baseline and the physical width of the steel strip centerline are calculated, including: Extract a set of paired {left edge points, right edge points} from the point set of the left edge in world coordinates and the point set of the right edge in world coordinates; Based on the set of paired {left edge points, right edge points}, calculate the set of widths and the set of deviations; The mean of the set of widths is calculated as the physical width of the centerline of the steel strip; The mean of the set of checks is calculated as the physical deviation of the steel strip centerline relative to the equipment baseline.

5. The image recognition and positioning method for automatic feeding of steel strips of varying thicknesses according to claim 4, characterized in that, If the confidence level of the left edge is less than a preset threshold but the confidence level of the right edge is greater than a preset threshold, the final deviation is calculated based on the point set of the right edge in world coordinates and the physical width of the steel strip centerline measured in the last time, including: Calculate the mean X-axis coordinates of the point set of the right edge in world coordinates to obtain the position of the right edge; The final deviation is obtained by subtracting half of the physical width of the steel strip centerline from the previously measured right edge position.

6. The image recognition and positioning method for automatic feeding of steel strips of varying thicknesses according to claim 4, characterized in that, If the confidence level of the right edge is less than a preset threshold but the confidence level of the left edge is greater than a preset threshold, the final deviation is calculated based on the point set of the left edge in world coordinates and the physical width of the steel strip centerline measured in the last time, including: Calculate the mean X-axis coordinates of the set of points on the left edge in world coordinates to obtain the position of the left edge; The final deviation is obtained by subtracting half of the physical width of the steel strip centerline from the previously measured left edge position.

7. The image recognition and positioning method for automatic feeding of steel strips of varying thicknesses according to claim 1, characterized in that, The final deviation and overall confidence level are input into the PLC controller to obtain control signals, including: The overall confidence level is multiplied by the proportional gain of the PLC controller to obtain the updated proportional gain.

8. An image recognition and positioning system for automatic feeding of steel strips of varying thicknesses, characterized in that, include: The raw steel strip image acquisition module is used to acquire raw image frames of steel strips captured by an industrial camera. A region of interest (ROI) localization module for steel strip images is used to quickly locate the ROI of the original steel strip image frame to obtain the left and right image block tensors. This includes: quickly locating the ROI of the original steel strip image frame to obtain left and right ROI rectangles; extracting left and right image blocks from the original steel strip image frame based on the left and right ROI rectangles; performing size modulation on the left and right image blocks based on bilinear interpolation to obtain size-normalized left and right image blocks; and normalizing the size-normalized left and right image blocks to obtain the left and right image block tensors. The edge feature regression module is used to input the tensors of the left and right image patches into the trained edge regression network to obtain the left edge parameters, right edge parameters, left edge confidence, and right edge confidence. The edge point set coordinate transformation module is used to sample the left ROI rectangle and the right ROI rectangle, and map the sampled points from the local coordinate system to the world coordinate system based on the left edge parameters, the right edge parameters and the homography matrix to obtain the point set of the left edge in the world coordinate system and the point set of the right edge in the world coordinate system. The steel strip geometry parameter calculation module is used to calculate the physical deviation of the steel strip centerline relative to the equipment baseline and the physical width of the steel strip centerline based on the point set of the left edge and the point set of the right edge in world coordinates. An edge confidence decision module is used to make robust decisions on the physical deviation of the steel strip centerline relative to the equipment baseline based on the left edge confidence and the right edge confidence to obtain the final deviation and a comprehensive confidence. This includes: taking the minimum of the left edge confidence and the right edge confidence as the comprehensive confidence; if both the left edge confidence and the right edge confidence are greater than a preset threshold, determining the physical deviation of the steel strip centerline relative to the equipment baseline as the final deviation; if the left edge confidence is less than the preset threshold but the right edge confidence is greater than the preset threshold, calculating the final deviation based on the point set of the right edge in world coordinates and the previously measured physical width of the steel strip centerline; and if the right edge confidence is less than the preset threshold but the left edge confidence is greater than the preset threshold, calculating the final deviation based on the point set of the left edge in world coordinates and the previously measured physical width of the steel strip centerline. The deviation control signal generation module is used to input the final deviation and the overall confidence level into the PLC controller to obtain the control signal.

Images