A method for automatically adding a covering agent into molten steel based on an image recognition algorithm

Abstract

The present application relates to the technical field of image recognition, and particularly relates to a method for automatically adding covering agent into molten steel based on an image recognition algorithm, which comprises the following steps: collecting molten steel liquid surface images by using a camera; improving the backbone and neck network of a YOLOv11 network, introducing a WIoU loss function, and constructing an FCW-YOLOv11s model; using a depth camera to capture molten steel liquid surface images in real time, inputting the collected images into the improved target detection model which has been trained, detecting, and obtaining a detection result including the coordinate information of each exposed part of the molten steel liquid surface and the area of the gray scale image; a computer judges the amount of covering agent according to the detection result, and transmits corresponding data to a mechanical arm, and the mechanical arm completes the operation of adding covering agent. The present application can realize real-time monitoring of the exposed condition of the molten steel liquid surface, and automatically add covering agent into the molten steel through the mechanical arm, thereby improving the rationality of the addition of covering agent.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and specifically to a method for automatically adding a covering agent to molten steel based on an image recognition algorithm. Background Technology

[0002] With the continuous optimization of steel production technology and the demand for intelligent transformation in the metallurgical industry, modern steel enterprises, in pursuit of efficient, high-quality, safe, low-cost, and green low-carbon production, find that traditional methods of manually adding covering agents are no longer sufficient to meet the current demands for high efficiency and high quality. Therefore, exploring a method that can achieve precise, uniform, and intelligent addition of covering agents is crucial.

[0003] After refining, the heat retention performance of the molten steel in the ladle is crucial for ensuring casting temperature and subsequent billet quality, and has therefore always been a focus of attention in the metallurgical industry. Traditionally, after the refining process, the addition of covering agent in actual production is generally done manually. However, due to today's automated and continuous production pace and limited platform space, manual addition of covering agent often tends to be excessive to ensure coverage, directly leading to waste. Furthermore, it is difficult to ensure uniform spreading of the covering agent during addition, easily resulting in localized accumulation on the molten steel surface while other areas are insufficiently covered or even exposed, leading to poor rationality in covering agent addition. This not only causes a decrease in molten steel temperature, affecting casting quality, but also increases the difficulty of subsequent quality control due to secondary oxidation of the molten steel. Summary of the Invention

[0004] To address the technical problem of poor rationality in the addition of covering agent, this invention proposes a method for automatically adding covering agent to molten steel based on an image recognition algorithm. This method enables a computer to monitor the exposed state of the molten steel surface in real time and uses a robotic arm to automatically add covering agent to the molten steel.

[0005] This invention provides a method for automatically adding a covering agent to molten steel based on an image recognition algorithm, the method comprising: Step 1: Use a camera to collect images of the molten steel surface, label each image, and then divide the dataset into a training set, a validation set, and a test set to form the initial dataset. Step 2: Based on the YOLOv11 network, improve the backbone and neck networks of the YOLOv11 network, introduce the WIoU loss function with dynamic non-monotonic focusing mechanism, and construct the FCW-YOLOv11s model to obtain the improved target detection model. Step 3: Input the initial dataset into the improved object detection model for training until the loss function WIoU converges, thus completing the training of the improved object detection model. Step 4: Calibrate the depth camera using Zhang Zhengyou's calibration method; Step 5: Based on the calibration of the depth camera, perform hand-eye calibration on the robotic arm to obtain the relationship between the robotic arm coordinate system and the camera coordinate system, thereby calculating the coordinates of the cover in the robotic arm coordinate system and obtaining the position and orientation of the depth camera in the robotic arm coordinate system. Step 6: Use a depth camera to capture images of the molten steel surface in real time. Input the captured images into the trained and improved target detection model for detection. Based on the position and orientation of the depth camera in the robot arm coordinate system, obtain the coordinate information of each exposed part of the molten steel surface and the detection results of the grayscale area. Step 7: The computer determines the amount of cover agent to be added based on the detection results and transmits the corresponding data to the robotic arm, which then completes the operation of adding the cover agent.

[0006] Optionally, the construction process of the improved target detection model includes: Based on the YOLOv11 network, the Fast Spatial Pyramid Pooling (SPPF) module in the backbone network of the YOLOv11 network is improved into a lighter feature weight sharing pyramid (FSP) module. The Upsample module in the neck network of the YOLOv11 network used for feature upsampling is improved into a module CSAS that facilitates inter-channel information interaction. The training loss function CIoU of YOLOv11 is improved into the WIoU loss function that introduces a dynamic non-monotonic focusing mechanism.

[0007] Optionally, the specific process of the lightweight feature weight sharing pyramid module FSP includes: The output of the previous layer is used as the input of the lightweight feature pyramid module FSP. The input feature map is processed by a convolution with a kernel of 1×1 and an output channel of C / 2 to obtain the feature map after channel compression, where C is the number of input channels. Using three 3×3 shared weight convolutions with dilation rates of 1, 3 and 5, the feature map after channel compression is first processed by the convolution with a dilation rate of 1. The output of the processed convolution is used as the input of the convolution with a dilation rate of 3, and the output of the convolution with a dilation rate of 3 is used as the input of the convolution with a dilation rate of 5 to obtain the preliminary feature map. The channel-compressed feature map is concatenated with the convolution outputs from step 2, which have dilation rates of 1, 3, and 5, to obtain the fused feature map. The obtained fused feature map is convolved using a 1×1 convolution kernel to adjust the number of channels, and then fused with the multi-scale features generated in step 3 to obtain a multi-scale feature map.

[0008] Optionally, the specific procedures of module CSAS include: The output of the previous layer is used as the input of the CSAS module. The Upsample operator is used to enlarge the height and width of the image to twice the original size. A depthwise separable convolution with a group number equal to the number of input channels and a kernel of 3×3 is used to perform spatial filtering on each input channel to obtain the enlarged feature map. The enlarged feature map is processed by channel rearrangement, the shape of the input tensor is reshaped, the reshaped tensor is transposed and reshaped back to its initial shape, thus realizing the channel order rearrangement operation and obtaining the rearranged feature map. The rearranged feature map is divided into four parts along the channel dimension: x1, x2, x3, and x4. Each part is cyclically shifted in a different direction: x1 is shifted positively along the height direction, x2 is shifted negatively along the height direction, x3 is shifted positively along the width direction, and x4 is shifted negatively along the width direction. The four shifted parts are then merged to obtain the shifted feature map. The shifted feature map is processed using a 1×1 convolution to obtain the inter-channel information interaction feature map.

[0009] Optionally, the expression for the WIoU loss function is: ; ; ; ; ; in, This represents the loss value for WIoU version 1. This is a penalty term for the distance attention weight, i.e., WIoU. This represents the IoU loss value. and These are the x and y coordinates of the predicted bounding box center point, respectively. and These are the x and y coordinates of the center point of the actual annotation box, respectively. , These are the width and height of the smallest bounding rectangle that can simultaneously enclose the predicted bounding box and the ground truth bounding box, respectively; the superscript * indicates a separation operation. and It is separated during computation and does not participate in gradient calculation; Indicates outlier degree, used to measure anchor frame quality; for The average value, It is the monotonic focusing coefficient; , All are hyperparameters. The non-monotonic focusing coefficient, This represents the loss value for WIoU version 3.

[0010] The present invention has the following beneficial effects: This invention discloses a method for automatically adding a covering agent to molten steel based on an image recognition algorithm. By improving the Fast Spatial Pyramid Pooling (SPPF) module in the backbone network of the YOLOv11 network to a lighter Feature Weight Sharing Pyramid (FSP) module, it significantly enhances the model's ability to capture multi-scale contextual information. While maintaining strong feature extraction capabilities, it reduces the number of model parameters, resulting in higher computational efficiency. Furthermore, this invention improves the Upsample module in the neck network of the YOLOv11 network used for feature upsampling to a CSAS module that facilitates inter-channel information interaction. Through channel rearrangement, spatial shifting, and point convolution operations, it transforms the input image, thereby promoting inter-channel information interaction and enhancing feature expressiveness. Finally, this invention improves the YOLOv11 training loss function CIoU to the WIoU loss function. By introducing a dynamic non-monotonic focusing mechanism, it enhances the model's focus on sample quality, reduces the weight of high-quality and low-quality samples, helps improve the accuracy of detection box judgments, accelerates model convergence, and gives the model better generalization ability and robustness. Therefore, this invention improves the accuracy of real-time detection of exposed molten steel surfaces while accelerating model convergence, giving the model higher generalization ability and robustness. Furthermore, by combining depth camera technology with a robotic arm, it enables the automatic addition of a covering agent to the molten steel, thereby improving the rationality of the covering agent addition. This invention has significant advantages over traditional manual addition of covering agents in practical applications. Attached Figure Description

[0011] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of a method for automatically adding a covering agent to molten steel based on an image recognition algorithm according to the present invention; Figure 2 This is a schematic diagram of the CSAS module of the present invention; Figure 3 This is a schematic diagram of the FCW-YOLOv11s convolutional neural network model of the present invention; Figure 4 This is a schematic diagram of the FSP module of the present invention; Figure 5 This is a schematic diagram of a real-world scenario illustrating a method for automatically adding a covering agent to molten steel based on an image recognition algorithm, according to the present invention. Detailed Implementation

[0013] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects of the technical solution proposed according to the present invention are described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0014] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0015] refer to Figure 1 This document illustrates the flowchart of some embodiments of a method for automatically adding a covering agent to molten steel based on an image recognition algorithm, according to the present invention. The method includes the following steps: Step 1: Use a camera to collect images of the molten steel surface, label each image, and then divide the dataset into a training set, a validation set, and a test set to form the initial dataset.

[0016] In some embodiments, a camera can be used to collect images of the molten steel surface, each image can be labeled with data, and the dataset can be divided into a training set, a validation set, and a test set in a ratio of 8:1:1 to form an initial dataset. Data augmentation processing can then be applied to the training set separately to obtain the augmented dataset.

[0017] The data augmentation methods in step 1 include: horizontal and vertical flipping, adjusting brightness and contrast, blurring, and sharpening.

[0018] Step 2: Based on the YOLOv11 network, improve the backbone and neck networks of the YOLOv11 network, introduce the WIoU loss function with dynamic non-monotonic focusing mechanism, and construct the FCW-YOLOv11s model to obtain the improved target detection model.

[0019] In some embodiments, based on the YOLOv11 network, the backbone and neck networks of the YOLOv11 network are improved, and a WIoU loss function with a dynamic non-monotonic focusing mechanism is introduced to construct an FCW-YOLOv11s model.

[0020] As an example, the specific construction process of the improved object detection model in step 2 is as follows: Based on the YOLOv11 network, firstly, the fast spatial pyramid pooling module SPPF in the backbone network of the YOLOv11 network is improved to a lighter feature weight shared pyramid module FSP (Feature Shared Pyramid). Secondly, the upsample module used for feature upsampling in the neck network of the YOLOv11 network is improved to a module CSAS (Channel Shuffling and Shift) which facilitates information interaction between channels. Finally, the training loss function CIoU of YOLOv11 is improved to the WIoU loss function which introduces a dynamic non-monotonic focusing mechanism.

[0021] like Figure 2 As shown, to improve the detection accuracy of exposed targets on the molten steel surface, based on the YOLOv11 network, the backbone and neck networks of the YOLOv11 network are improved, and a WIoU loss function with a dynamic non-monotonic focusing mechanism is introduced to construct an FCW-YOLOv11s model; specifically, as... Figure 3 As shown, the Fast Spatial Pyramid Pooling (SPPF) module in the YOLOv11 backbone network is improved to a lighter Feature Weight Sharing Pyramid (FSP) module. Through parameter sharing and the formation of multi-scale receptive fields, this significantly enhances the model's ability to capture multi-scale contextual information. Furthermore, while maintaining strong feature extraction capabilities, it also boasts high computational efficiency, thereby improving the model's detection accuracy for exposed targets on molten steel surfaces. The backbone network undertakes the core task of extracting depth features from the molten steel surface image. Through a structure composed of convolutional layers, pooling layers, and activation functions, it progressively transforms the original input image (a 640×640 three-channel RGB image) into a high-dimensional feature map rich in semantic information, detailed features, and contextual relationships, providing a foundation for subsequent target recognition. The neck network is the key module connecting the backbone network and the head. It further integrates and enhances the multi-scale features generated by the backbone network by fusing top-down and bottom-up bidirectional paths. This structure determines the model's ability to detect targets of different sizes. To improve the information interaction capability between feature channels in the neck network of the YOLOv11s model, such as Figure 4As shown, by improving the Upsample module for feature upsampling in the neck network of the YOLOv11 network to the CSAS module, which facilitates inter-channel information interaction, the input image is transformed through channel rearrangement, spatial shifting, and point convolution operations. This allows for thorough mixing of information between feature map channels, thereby enhancing the model's ability to capture feature map details and promoting inter-channel information interaction while strengthening feature representation. To improve the model's robustness and convergence speed, the YOLOv11 training loss function CIoU is improved to WIoU. By introducing a dynamic non-monotonic focusing mechanism, the model pays more attention to sample quality, reducing the weights of high-quality and low-quality samples. This helps improve the accuracy of bounding box judgments and accelerates model convergence, giving the model better generalization ability and robustness, thus improving the overall detection performance. A practical scenario for a method of automatically adding a covering agent to molten steel based on an image recognition algorithm can be described as follows. Figure 5 As shown.

[0022] The lightweight Feature Pyramid Module (FSP) is used for multi-scale extraction of molten steel surface features. The specific process includes the following steps: The first step is to use the output of the previous layer as the input of the lightweight feature pyramid module FSP, and then use a convolution with a kernel of 1×1 and an output channel of C / 2 to process the input feature map, so as to obtain the feature map after channel compression.

[0023] Where C represents the number of input channels.

[0024] The second step involves using three 3×3 shared weight convolutions with dilation rates of 1, 3, and 5. The feature map after channel compression is first processed by the convolution with a dilation rate of 1. Then, the processed output is used as the input to the convolution with a dilation rate of 3. Finally, the output of the convolution with a dilation rate of 3 is used as the input to the convolution with a dilation rate of 5. This effectively captures multi-scale contextual information and yields a preliminary feature map.

[0025] The third step is to concatenate the channel-compressed feature map with the convolution outputs from step 2, which have dilation rates of 1, 3, and 5, to fuse features at different scales and obtain a fused feature map.

[0026] The fourth step involves convolving the obtained fused feature map with a 1×1 convolution kernel to adjust the number of channels, and then fusing the multi-scale features generated in step 3 to obtain a multi-scale feature map.

[0027] The CSAS module, which facilitates information exchange between channels, is used to promote information exchange between channels representing molten steel surface features. The specific process may include the following steps: The first step is to use the output of the previous layer as the input of the CSAS module. First, the Upsample operator is used to enlarge the height and width of the image to twice their original size. Then, a depthwise separable convolution with a group number equal to the number of input channels and a kernel of 3×3 is used to perform spatial filtering independently on each input channel, reducing the amount of computation while preserving spatial features, and obtaining the enlarged feature map.

[0028] The second step, channel rearrangement processing of the enlarged feature map, may include the following steps: First, the shape of the input tensor is reshaped from batchsize, num_channel, height, and width to batchsize, group, per_group_channel, height, and width. Then, the reshaped tensor is transposed and reshaped back to its initial shape to achieve channel order rearrangement and obtain the rearranged feature map.

[0029] The third step is to divide the rearranged feature map into four parts along the channel dimension, namely x1, x2, x3, and x4. For each part, perform a cyclic shift in different directions: part x1 is shifted positively along the height direction, part x2 is shifted negatively along the height direction, part x3 is shifted positively along the width direction, and part x4 is shifted negatively along the width direction. Finally, merge the four shifted parts to restore the original number of channels, enhance the global perception capability of the features, and obtain the shifted feature map.

[0030] The fourth step is to process the shifted feature map using a 1×1 convolution to promote the fusion of feature information between feature map channels and obtain an inter-channel information interaction feature map.

[0031] The expression for the WIoU loss function of the dynamic non-monotonic focusing mechanism, used to improve the accuracy of detection boxes and accelerate model convergence, is as follows: ; ; ; ; ; in, This represents the loss value for WIoU version 1. This is a penalty term for the distance attention weight, i.e., WIoU. This represents the IoU loss value. and These are the x and y coordinates of the predicted bounding box center point, respectively. and These are the x and y coordinates of the center point of the actual annotation box, respectively. , These are the width and height of the smallest bounding rectangle that can simultaneously enclose both the predicted and ground truth bounding boxes, respectively. The superscript * indicates a separation operation. and It is separated during computation and does not participate in gradient calculation. This indicates the outlier degree and is used to measure the quality of the anchor frame. for The average value, It is the monotonic focusing coefficient. , All are hyperparameters. The non-monotonic focusing coefficient, This represents the loss value for WIoU version 3.

[0032] Step 3: Input the initial dataset into the improved object detection model for training until the loss function WIoU converges, thus completing the training of the improved object detection model.

[0033] In some embodiments, the initial dataset can be input into the improved object detection model for training until the loss function WIoU converges, thus completing the training of the improved object detection model.

[0034] Step 4: Use Zhang Zhengyou's calibration method to calibrate the depth camera.

[0035] In some embodiments, the Zhang Zhengyou calibration method can be used to calibrate the depth camera, and the obtained depth camera intrinsic parameter matrix, tangential distortion, radial distortion, baseline, and relative positions of the left and right cameras are obtained.

[0036] Step 5: Based on the calibration of the depth camera, perform hand-eye calibration on the robotic arm to obtain the relationship between the robotic arm coordinate system and the camera coordinate system, thereby calculating the coordinates of the cover in the robotic arm coordinate system and obtaining the position and orientation of the depth camera in the robotic arm coordinate system.

[0037] In some embodiments, the depth camera and the robotic arm can be installed and calibrated using an eye-to-hand method, taking into account the actual scenario. The robotic arm can be calibrated by hand and eye to obtain the relationship between the robotic arm coordinate system and the camera coordinate system. This allows the coordinates of the cover in the robotic arm coordinate system to be calculated, and finally the position and orientation of the depth camera in the robotic arm coordinate system to be obtained.

[0038] Step 6: Use a depth camera to capture images of the molten steel surface in real time. Input the captured images into the trained and improved target detection model for detection. Based on the position and orientation of the depth camera in the robot arm coordinate system, obtain the detection results including the coordinate information of each exposed part of the molten steel surface and the grayscale area.

[0039] In some embodiments, a depth camera can be used to capture images of the molten steel surface in real time. The captured images are then input into a trained and improved target detection model for detection, thereby obtaining the coordinate information and grayscale area of ​​each exposed part of the molten steel surface.

[0040] Step 7: The computer determines the amount of cover agent to be added based on the detection results and transmits the corresponding data to the robotic arm, which then completes the operation of adding the cover agent.

[0041] In some embodiments, the computer determines the amount of cover agent to be added based on the detection results and transmits the corresponding data to the robotic arm, which then performs the operation of adding the cover agent.

[0042] In summary, this invention improves the Fast Spatial Pyramid Pooling (SPPF) module in the backbone network of the YOLOv11 network by replacing it with a lighter Feature Weight Sharing Pyramid (FSP) module. This significantly enhances the model's ability to capture multi-scale contextual information and reduces the number of model parameters while maintaining strong feature extraction capabilities, resulting in higher computational efficiency. Furthermore, by improving the Upsample module in the neck network of the YOLOv11 network used for feature upsampling, it replaces it with the CSAS module, which facilitates inter-channel information interaction. Through channel rearrangement, spatial shifting, and point convolution operations, the input image is transformed, thereby promoting inter-channel information interaction and enhancing feature expressiveness. Finally, this invention improves the YOLOv11 training loss function CIoU to the WIoU loss function. By introducing a dynamic non-monotonic focusing mechanism, it enhances the model's focus on sample quality, reduces the weights of high-quality and low-quality samples, helps improve the accuracy of bounding box judgments, accelerates model convergence, and gives the model better generalization ability and robustness. Therefore, this invention improves the accuracy of real-time detection of exposed molten steel surfaces while accelerating model convergence, giving the model higher generalization ability and robustness. Furthermore, by combining depth camera technology with a robotic arm, it enables the automatic addition of a covering agent to the molten steel, thereby improving the rationality of the covering agent addition. This invention has significant advantages over traditional manual addition of covering agents in practical applications.

[0043] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for automatically adding a covering agent to molten steel based on an image recognition algorithm, characterized in that, Includes the following steps: Step 1: Use a camera to collect images of the molten steel surface, label each image, and then divide the dataset into a training set, a validation set, and a test set to form the initial dataset. Step 2: Based on the YOLOv11 network, improve the backbone and neck networks of the YOLOv11 network, introduce the WIoU loss function with dynamic non-monotonic focusing mechanism, and construct the FCW-YOLOv11s model to obtain the improved target detection model. Step 3: Input the initial dataset into the improved object detection model for training until the loss function WIoU converges, thus completing the training of the improved object detection model. Step 4: Calibrate the depth camera using Zhang Zhengyou's calibration method; Step 5: Based on the calibration of the depth camera, perform hand-eye calibration on the robotic arm to obtain the relationship between the robotic arm coordinate system and the camera coordinate system, thereby calculating the coordinates of the cover in the robotic arm coordinate system and obtaining the position and orientation of the depth camera in the robotic arm coordinate system. Step 6: Use a depth camera to capture images of the molten steel surface in real time. Input the captured images into the trained and improved target detection model for detection. Based on the position and orientation of the depth camera in the robot arm coordinate system, obtain the coordinate information of each exposed part of the molten steel surface and the detection results of the grayscale area. Step 7: The computer determines the amount of cover agent to be added based on the detection results and transmits the corresponding data to the robotic arm, which then completes the operation of adding the cover agent.

2. The method for automatically adding a covering agent to molten steel based on an image recognition algorithm according to claim 1, characterized in that, The construction process of the improved target detection model includes: Based on the YOLOv11 network, the Fast Spatial Pyramid Pooling (SPPF) module in the backbone network of the YOLOv11 network is improved into a lighter feature weight sharing pyramid (FSP) module. The Upsample module in the neck network of the YOLOv11 network used for feature upsampling is improved into a module CSAS that facilitates inter-channel information interaction. The training loss function CIoU of YOLOv11 is improved into the WIoU loss function that introduces a dynamic non-monotonic focusing mechanism.

3. The method for automatically adding a covering agent to molten steel based on an image recognition algorithm according to claim 2, characterized in that, The specific process of the lightweight feature weight sharing pyramid module (FSP) includes: The output of the previous layer is used as the input of the lightweight feature pyramid module FSP. The input feature map is processed by a convolution with a kernel of 1×1 and an output channel of C / 2 to obtain the feature map after channel compression, where C is the number of input channels. Using three 3×3 shared weight convolutions with dilation rates of 1, 3 and 5, the feature map after channel compression is first processed by the convolution with a dilation rate of 1. The output of the processed convolution is used as the input of the convolution with a dilation rate of 3, and the output of the convolution with a dilation rate of 3 is used as the input of the convolution with a dilation rate of 5 to obtain the preliminary feature map. The channel-compressed feature map is concatenated with the convolution outputs from step 2, which have dilation rates of 1, 3, and 5, to obtain the fused feature map. The obtained fused feature map is convolved using a 1×1 convolution kernel to adjust the number of channels, and then fused with the multi-scale features generated in step 3 to obtain a multi-scale feature map.

4. The method for automatically adding a covering agent to molten steel based on an image recognition algorithm according to claim 2, characterized in that, The specific process of the CSAS module includes: The output of the previous layer is used as the input of the CSAS module. The Upsample operator is used to enlarge the height and width of the image to twice the original size. A depthwise separable convolution with a group number equal to the number of input channels and a kernel of 3×3 is used to perform spatial filtering on each input channel to obtain the enlarged feature map. The enlarged feature map is processed by channel rearrangement, the shape of the input tensor is reshaped, the reshaped tensor is transposed and reshaped back to its initial shape, thus realizing the channel order rearrangement operation and obtaining the rearranged feature map. The rearranged feature map is divided into four parts along the channel dimension: x1, x2, x3, and x4. Each part is cyclically shifted in a different direction: x1 is shifted positively along the height direction, x2 is shifted negatively along the height direction, x3 is shifted positively along the width direction, and x4 is shifted negatively along the width direction. The four shifted parts are then merged to obtain the shifted feature map. The shifted feature map is processed using a 1×1 convolution to obtain the inter-channel information interaction feature map.

5. A method for automatically adding a covering agent to molten steel based on an image recognition algorithm according to claim 1, characterized in that, The expression for the WIoU loss function is: ； ； ； ； ； in, This represents the loss value for WIoU version 1. This is a penalty term for the distance attention weight, i.e., WIoU. This represents the IoU loss value. and These are the x and y coordinates of the predicted bounding box center point, respectively. and These are the x and y coordinates of the center point of the actual annotation box, respectively. , These are the width and height of the smallest bounding rectangle that can simultaneously enclose the predicted bounding box and the ground truth bounding box, respectively; the superscript * indicates a separation operation. and It is separated during computation and does not participate in gradient calculation; Indicates outlier degree, used to measure anchor frame quality; for The average value, It is the monotonic focusing coefficient; , All are hyperparameters. The non-monotonic focusing coefficient, This represents the loss value for WIoU version 3.

Images