A molten iron light image enhancement method and an intrusion detection method in a steel plant

Abstract

The application discloses a molten iron light image enhancement method and an intrusion detection method in a steel plant, and comprises the following steps: obtaining a molten iron light image; based on a generation network, taking the obtained molten iron light image as a target of the generation network, and optimizing the generation network; generating a molten iron light feature map by using the optimized generation network; attaching the generated molten iron light feature map to the regions of the steel plant image to realize the enhancement of the steel plant image; based on a YOLOv5 model, taking all the enhanced images B as a training set to train the YOLOv5 model; and performing target detection by using the trained YOLOv5 model. The application greatly expands the training data of the YOLOv5 model by performing image enhancement on the image of the region with the molten iron light in the steel plant, provides rich and diverse training data for the YOLOv5 model, and greatly improves the accuracy of target detection of the YOLOv5 model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of visual recognition intrusion detection in steel plants, and particularly relates to a molten iron light image enhancement method and an intrusion detection method in a steel plant. BACKGROUND

[0002] In a steel plant, due to the particularity of work, there are many dangerous areas, such as ladle car, slag pot car, and travelling car operation area. When other personnel or vehicles enter these dangerous areas, safety accidents are prone to occur. For safety production, it is necessary to set up visual recognition-based intrusion detection in these dangerous areas, that is, to identify whether other personnel or vehicles have intruded into the dangerous area. If other personnel or vehicles have intruded into the dangerous area, timely alarm or interruption of work is facilitated. Unlike intrusion detection in other environments, there is high-brightness molten iron light in a steel plant. In the ladle car, slag pot car operation area or ladle area, due to the dynamic properties of the movement of the ladle car, slag pot car and ladle area operation, molten iron light frequently appears at different positions in the related area, changing the area environment, which is diverse and complex. The molten iron light appears frequently, but the light area is diverse, and the light speed is fast, making it difficult to obtain a steel plant picture with molten iron light, the model training set sample is small, the detection accuracy is low, and the detection difficulty is great. If the molten iron light is not identified in the intrusion detection of other personnel or vehicles, it will bring safety hazards. SUMMARY

[0003] The purpose of the present application is to provide a molten iron light image enhancement method in a steel plant.

[0004] The technical solution for achieving the purpose of the present application is: a molten iron light image enhancement method in a steel plant, comprising the following steps:

[0005] S1. Take a picture with molten iron light, set the picture with molten iron light as picture A, and crop the place affected by light in picture A to obtain a molten iron light image;

[0006] S2. Based on the generation network, the molten iron light image obtained in step S1 is taken as the target of the generation network, and the generation network is optimized. The step of optimizing the generation network comprises:

[0007] The random vector is input into the generation network to be optimized, the generation process graph is generated, the loss value of the process graph and the molten iron light image is calculated, and then the optimization algorithm is used to update the generation network to be optimized to obtain the optimized generation network;

[0008] The above optimization steps are continuously repeated more than once; when the optimization step is performed for the first time, the initial generation network before optimization is taken as the model to be optimized; in each subsequent optimization step, the optimized generation network obtained in the previous optimization step is taken as the model to be optimized;

[0009] S3. Generating the molten iron light feature map using the final optimized generation network obtained in step S2;

[0010] S4. Obtaining an image in a steel plant, selecting more than one region on the image in the steel plant, and attaching the molten iron light feature map generated in step S3 to the regions on the image in the steel plant to achieve enhancement of the image in the steel plant.

[0011] Further, when the size of the cropped molten iron light map is inconsistent with the two-dimensional size of the process map, the cropped molten iron light map is scaled by the bilinear interpolation method before the loss value of the process map and the molten iron light map is calculated, so that the size of the molten iron light map is consistent with the two-dimensional size of the process map input to the generation network. The molten iron light map as the target of the generation network usually needs to be consistent with the two-dimensional size of the process map generated by the generation network. In order to make the molten iron light map as the target of the generation network consistent with the two-dimensional size of the process map, the place in the picture A affected by the light is cropped to obtain the molten iron light map, and the molten iron light map consistent with the two-dimensional size of the process map can be cropped. When the size of the cropped molten iron light map is inconsistent with the two-dimensional size of the process map, the cropped molten iron light map can also be scaled by the bilinear interpolation method, so that the size of the molten iron light map is consistent with the two-dimensional size of the process map.

[0012] Further, in step S2, the loss value of the process map and the molten iron light map is calculated according to the formula , wherein is the loss value of the process map and the molten iron light map, is the molten iron light map, is the generation network, is a random vector input to the generation network, is the batch size of the input random vector, is a positive integer.

[0013] Further, in step S2, the optimization algorithm is any one of the random gradient descent method, the Nesterov gradient acceleration method, and the Adam adaptive time estimation method.

[0014] Further, in step S3, the final optimized generation network is used to generate the molten iron light feature map, and the input of the final optimized generation network is a random vector.

[0015] Further, in step S4, the molten iron light feature map generated in step S3 is attached to the regions of the steel plant image by a linear blending algorithm or a Poisson blending algorithm.

[0016] Further, the molten iron light feature map generated in step S3 is attached to the regions of the steel plant image by a linear blending algorithm, and the formula of the linear blending algorithm is , wherein is the enhanced image, is the steel plant image, is the molten iron light feature map, is the superposition weight, The value of is between 0 and 1. Compared with the Poisson blending algorithm, the linear blending algorithm is more commonly used.

[0017] Further, in step S2, the initially set generation network is composed of 5 deconvolution network layers, wherein the first deconvolution input channel number is 100, the output channel number is 512, and the convolution step length is 1; the second deconvolution input channel number is 512, the output channel number is 256, and the convolution step length is 2; the third deconvolution input channel number is 256, the output channel number is 128, and the convolution step length is 2; the fourth deconvolution input channel number is 128, the output channel number is 64, and the convolution step length is 2; and the fifth deconvolution input channel number is 64, the output channel number is 3, and the convolution step length is 2.

[0018] The second object of the present application is to provide an intrusion detection method.

[0019] The technical solution for achieving the second object of the present application is as follows: an intrusion detection method for detecting intrusion in a region with molten iron light in a steel plant, comprising the following steps:

[0020] J1. Batch acquisition of images of regions with molten iron light in a steel plant, and the images of the regions with molten iron light in the steel plant are images B; all the images B are image enhanced by the steel plant molten iron light image enhancement method of the first object of the present application, wherein in step S4 of the steel plant molten iron light image enhancement method of the present application, the selection method of selecting more than one region on the steel plant image is random selection;

[0021] J2. Based on the YOLOv5 model, all the enhanced images B are input into the YOLOv5 model as a training set to train the YOLOv5 model;

[0022] J3. Real-time acquisition of a detection image of a to-be-detected area in a steel plant, the real-time acquired detection image of the to-be-detected area in the steel plant being image C, target detection of the image C is performed by the YOLOv5 model trained in step J2; if a target is detected, the positioning of the target in the image C is acquired, and it is judged whether the target is located in the divided dangerous area, so as to realize intrusion detection of the to-be-detected area in the steel plant.

[0023] Further, in step S4, four areas are selected on the image in the steel plant. Not all places in the steel plant are irradiated by molten iron light, that is, not the whole image in the steel plant has molten iron light. According to observation, the probability of the four areas of the image in the steel plant having molten iron light is relatively large. After the four areas are selected on the image in the steel plant, the detection accuracy of the trained YOLOv5 model is higher when performing target detection.

[0024] Further, in step J2, the specific steps of training the YOLOv5 model include: iteratively inputting the training set data into the model for forward propagation, calculating the loss function of the YOLOv5 model, and continuously adjusting the weights of the YOLOv5 model through back propagation. By continuously inputting, calculating, and updating the parameters of the YOLOv5 model, the error of the target detection is finally minimized, and the training of the YOLOv5 model is completed.

[0025] Further, in step J3, the ray method, the odd-even rule or the non-zero rule is used to judge whether the target is located in the divided dangerous area.

[0026] Further, the intrusion detection method is realized based on an intrusion detection system, the intrusion detection system comprising an intrusion detection camera, a Python intrusion detection system, a C# data transmission interface, a PLC data processing end, and a WinCC display unit, the C# data transmission interface and the Python intrusion detection system being automatically started after being started;

[0027] After being started, the Python intrusion detection system periodically sends a normal operation signal to the C# data transmission interface. When the C# data transmission interface does not timely receive the normal operation signal sent by the Python intrusion detection system, it is considered that the Python intrusion detection system is abnormal, and the C# data transmission interface sends a signal of the abnormal Python intrusion detection system to the PLC data processing end, which is sent to the WinCC display unit for display;

[0028] The C# data transmission interface periodically sends a normal operation signal to the PLC data processing end, when the PLC data processing end does not receive the normal operation signal sent by the C# data transmission interface in time, it is considered that the C# data transmission interface is abnormal, and the PLC data processing end sends a signal of the abnormal C# data transmission interface to the WinCC display unit for display. In the intrusion detection system, the C# data transmission interface performs heartbeat detection on the Python intrusion detection, and at the same time, the PLC data processing end performs heartbeat detection on the C# data transmission interface, so that the abnormality of the intrusion detection system can be found in time, and the safety of the on-site staff and the production vehicles is effectively improved.

[0029] The molten iron light image enhancement method in the steel plant of the present application is based on a generation network. After the generation network is input with a random vector more than once, the loss value is calculated, and the optimization is updated, the generation network can be optimized, the deep learning is completed, and then the generation network after deep learning can generate accurate molten iron light feature maps. After the molten iron light feature maps are attached to the images in the steel plant, the image enhancement of the images in the steel plant can be realized. In this way, when the image data with molten iron light in the training set is insufficient, and the detection accuracy of the intrusion detection module is low during the subsequent intrusion detection process, the molten iron light feature maps can be attached to the images in the steel plant by the molten iron light image enhancement method in the steel plant of the present application to enhance the images in the steel plant, to generate image data with molten iron light, and then to expand the image data with molten iron light in the training set and improve the detection accuracy of the intrusion detection module.

[0030] The intrusion detection method of the present application enhances the images of the area with molten iron light in the steel plant by the molten iron light image enhancement method in the steel plant of the present application, and in the enhancement process, the selection method of more than one area on the image in the steel plant in step S4 of the molten iron light image enhancement method in the steel plant of the present application is random selection. The randomness of random selection makes the enhanced image B complex and varied. After all the enhanced images B are used as the training set of the YOLOv5 model, the training data of the YOLOv5 model is greatly expanded, and rich and diverse training data is provided for the YOLOv5 model. Under the training of rich and diverse training data, the accuracy of target detection of the YOLOv5 model is greatly improved, and the reliability of intrusion detection is improved, avoiding the situation that other personnel or vehicles are not recognized during intrusion detection due to the small training set and low detection accuracy, and greatly improving the production safety. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is a flowchart of the molten iron light image enhancement method in the steel plant of the present application;

[0032] Figure 2 is an example picture of the present application for taking a picture with molten iron light in a steel plant;

[0033] Figure 3 is an example picture of the present application for taking a picture with molten iron light in a steel plant;

[0034] Figure 4 is an example picture of the present application for taking a picture with molten iron light in a steel plant;

[0035] Figure 5 is an example picture of the present application for taking a picture with molten iron light in a steel plant;

[0036] Figure 6 is a flow chart of the present application for intrusion detection. DETAILED DESCRIPTION

[0037] The preferred embodiments of the present application for enhancing molten iron light image in a steel plant and intrusion detection method will be described in detail below with reference to the accompanying drawings:

[0038] A method for enhancing molten iron light image in a steel plant, as shown in Figure 1 , includes the following steps:

[0039] S1. Take a picture with molten iron light, set the picture with molten iron light as picture A, and the picture A is as shown in Figure 2 , cut out the place affected by light in the picture A to get a molten iron light image;

[0040] S2. Based on the generation network, take the molten iron light image obtained in step S1 as the target of the generation network, and optimize the generation network; the step of optimizing the generation network includes:

[0041] input a random vector into the generation network to be optimized, generate a process graph, calculate the loss value of the process graph and the molten iron light image; after calculating the loss value of the process graph and the molten iron light image, update the generation network to be optimized by the stochastic gradient descent method to obtain an optimized generation network;

[0042] The above optimization step is repeated continuously more than once; when the optimization step is performed for the first time, the initial generation network before optimization is used as the model to be optimized; in each subsequent optimization step, the optimized generation network obtained in the previous optimization step is used as the model to be optimized;

[0043] S3. Input a random vector into the generation network optimized in step S2, and generate a molten iron light feature map using the final optimized generation network obtained in step S2, as shown in Figure 3 ;

[0044] S4. Obtain an image in a steel plant, such as Figure 4 As shown in the figure, select more than one region on the image in the steel plant, and attach the molten iron light feature map generated in step S3 to the regions on the image in the steel plant by linear mixing algorithm, as shown in the figure, to achieve the enhancement of the image in the steel plant. Figure 5

[0045] The molten iron light image enhancement method in the steel plant of the present application generates a molten iron light feature map by using a generation network. The generation network initially set can be but is not limited to consisting of 5 deconvolution network layers, wherein the first deconvolution input channel number is 100, the output channel number is 512, and the convolution step length is 1; the second deconvolution input channel number is 512, the output channel number is 256, and the convolution step length is 2; the third deconvolution input channel number is 256, the output channel number is 128, and the convolution step length is 2; the fourth deconvolution input channel number is 128, the output channel number is 64, and the convolution step length is 2; and the fifth deconvolution input channel number is 64, the output channel number is 3, and the convolution step length is 2.

[0046] After the generation network is initially set, the generation network needs to be optimized, and after optimization, an accurate molten iron light feature map can be generated. When optimizing the generation network, a random vector is input to the generation network to be optimized. The size of the elements in the random vector can not be limited to between 0 and 1, and the dimension can be but is not limited to 100*1. When the dimension of the random vector is 100*1, the 64*64*3 process map is obtained after the random vector passes through the 5 deconvolution network layers of the generation network, and the two-dimensional size of the process map is 64*64.

[0047] In the process of optimizing the generation network, after obtaining the process map, the loss value of the process map and the molten iron light image needs to be calculated. In order to calculate the loss value of the process map and the molten iron light image, the molten iron light image as the target of the generation network usually needs to be consistent with the two-dimensional size of the process map generated by the generation network. In order to make the two-dimensional size of the molten iron light image as the target of the generation network consistent with the process map, the part of the picture A affected by the light is cropped to obtain the molten iron light image, and the molten iron light image consistent with the two-dimensional size of the process map can be cropped. When the size of the cropped molten iron light image is inconsistent with the two-dimensional size of the process map, the bilinear interpolation method can also be used to scale the cropped molten iron light image before calculating the loss value of the process map and the molten iron light image, so that the size of the molten iron light image is consistent with the two-dimensional size of the process map. For example, the two-dimensional size of the process map is 64*64, and the size of the molten iron light image is 64*64 whether it is cropped or scaled.

[0048] ​The loss value of the process graph and the molten iron light graph is calculated in the molten iron light image enhancement method in the steel plant , wherein is the loss value of the process graph and the molten iron light graph, is the molten iron light graph, is the generation network, is a random vector input to the generation network, is the batch size of the input random vector, is a positive integer.

[0049] After the loss value of the process graph and the molten iron light graph is calculated, the generation network to be optimized is updated by an optimization algorithm in the molten iron light image enhancement method in the steel plant. The optimization algorithm for updating the generation network to be optimized can be any one of the random gradient descent method, the Nesterov gradient acceleration method, and the Adam adaptive time estimation method.

[0050] After the input random vector, the loss value calculation, and the optimization update operation are repeated once or more times, the generation network is optimized, and the optimized generation network has completed deep learning and can generate a relatively accurate molten iron light feature map in the molten iron light image enhancement method in the steel plant. When the optimized generation network completes deep learning and generates a molten iron light feature map, its input can be set artificially or randomly. To simplify the operation, preferably, in step S3, when the molten iron light feature map is generated by using the finally optimized generation network in step S2, the input of the finally optimized generation network is a random vector, and the random vector is input by the program.

[0051] After the molten iron light feature map is obtained, it is attached to the image in the steel plant, and the image enhancement of the image in the steel plant can be realized in the molten iron light image enhancement method in the steel plant. When the molten iron light feature map is attached to the image in the steel plant, in addition to the linear mixing algorithm, the Poisson fusion algorithm can also be used to attach the molten iron light feature map generated in step S3 to the regions of the image in the steel plant.

[0052] Preferably, the linear mixing algorithm is used to attach the molten iron light feature map generated in step S3 to the regions of the image in the steel plant in the molten iron light image enhancement method in the steel plant, and the formula of the linear mixing algorithm is , wherein is the enhanced image, is the image in the steel plant, is the molten iron light feature map, is the superposition weight, The value of is between 0 and 1. Compared with the Poisson fusion algorithm, the linear mixing algorithm is more commonly used and is simpler to calculate.

[0053] The molten iron light image enhancement method in the steel plant based on the generation network can be optimized through the operation of inputting a random vector more than once, calculating a loss value, and updating optimization of the generation network, and deep learning is completed. Then, the generation network after deep learning can generate an accurate molten iron light feature map. After the molten iron light feature map is attached to the image in the steel plant, the image enhancement of the image in the steel plant can be realized. In this way, when the image data with molten iron light in the training set is insufficient, and the detection accuracy of the intrusion detection module is low, the molten iron light feature map can be attached to the image in the steel plant by the molten iron light image enhancement method in the steel plant to enhance the image in the steel plant, so as to generate image data with molten iron light, and then the image data with molten iron light in the training set can be expanded, and the detection accuracy of the intrusion detection module can be improved.

[0054] The application further provides an intrusion detection method. An intrusion detection method for detecting intrusion in a region with molten iron light in a steel plant, as shown in the following steps: Figure 6

[0055] J1. Batch acquisition of images of regions with molten iron light in the steel plant, the images of the regions with molten iron light in the steel plant being image B; image enhancement of all the images B by the molten iron light image enhancement method in the steel plant, wherein the selection of more than one region on the image in the steel plant in step S4 of the molten iron light image enhancement method in the steel plant is random selection;

[0056] J2. Based on the YOLOv5 model, all the enhanced images B are input into the YOLOv5 model as a training set to train the YOLOv5 model;

[0057] J3. Real-time acquisition of detection images of a region to be detected in the steel plant, the real-time acquisition of the detection images of the region to be detected in the steel plant being image C, target detection of the image C by the YOLOv5 model trained in step J2; if a target is detected, the positioning of the target in the image C is acquired, and whether the target is located in the divided dangerous region is judged to realize the intrusion detection of the region to be detected in the steel plant.

[0058] ​The application discloses an intrusion detection method, which adopts a YOLOv5 model for target detection, and the YOLOv5 model needs to be trained and optimized before target detection, so as to reduce the error of target detection. The accuracy of the YOLOv5 model detection is closely related to the diversity of the data in the training set, and the training set of the YOLOv5 model detection is composed of all the image B obtained by the image increasing through the molten iron light image enhancement method in a steel plant.

[0059] In the step S4, four regions are selected on the image in the steel plant. Not all places in the steel plant are irradiated by the molten iron light, that is to say, not the whole image in the steel plant has the molten iron light. According to observation, the probability of the four regions of the image in the steel plant having the molten iron light is relatively large. After the four regions are selected on the image in the steel plant, the trained YOLOv5 model has higher detection accuracy when performing target detection.

[0060] In the step J2, the specific steps of training the YOLOv5 model include: iteratively inputting the training set data into the model for forward propagation, calculating the loss function of the YOLOv5 model, and continuously adjusting the weight of the YOLOv5 model through back propagation. Through continuous input, calculation and update of the parameters of the YOLOv5 model, the error of target detection is finally minimized, and the training of the YOLOv5 model is completed.

[0061] In the step J3, the ray method, the parity rule or the non-zero rule are used to judge whether the target is located in the divided dangerous region. The ray method specifically refers to that a ray is made in a set direction, the number of intersection points of the ray and the dangerous region is calculated, and whether the target is located in the dangerous region is judged according to the parity of the number of intersection points. For example, when the number of intersection points is odd, the target is located in the dangerous region, and intrusion occurs. When the number of intersection points is even, the target is located outside the dangerous region, and intrusion does not occur. The parity rule specifically refers to that the rotation direction of the path of the target passing through each corner of the dangerous region is calculated, and the clockwise rotation is a positive value and the counterclockwise rotation is a negative value. If the sum of the absolute values of the two is 2pi, the target is in the dangerous region, and intrusion occurs. Otherwise, the target is outside the dangerous region, and intrusion does not occur. The non-zero rule specifically refers to that the "turn number" of the target to each side of the dangerous region is calculated. If the "turn number" is not 0, the target is in the dangerous region, and intrusion occurs. Otherwise, the target is outside the dangerous region, and intrusion does not occur.

[0062] The application discloses an intrusion detection method, which adopts a YOLOv5 model for target detection, and the YOLOv5 model needs to be trained and optimized before target detection, so as to reduce the error of target detection. The accuracy of the YOLOv5 model detection is closely related to the diversity of the data in the training set, and the training set of the YOLOv5 model detection is composed of all the image B obtained by the image increasing through the molten iron light image enhancement method in a steel plant.

[0063] After starting, the Python intrusion detection system periodically sends a normal operation signal to the C# data transmission interface, and when the C# data transmission interface does not receive the normal operation signal sent by the Python intrusion detection system in time, it is considered that the Python intrusion detection system is abnormal, and the C# data transmission interface sends a signal of the abnormality of the Python intrusion detection system to the PLC data processing end, which is sent to the WinCC display unit by the PLC data processing end for display.

[0064] The C# data transmission interface periodically sends a normal operation signal to the PLC data processing end, and when the PLC data processing end does not receive the normal operation signal sent by the C# data transmission interface in time, it is considered that the C# data transmission interface is abnormal, and the PLC data processing end sends a signal of the abnormality of the C# data transmission interface to the WinCC display unit for display.

[0065] The intrusion detection method of the present application, in the intrusion detection system, the C# data transmission interface carries out heartbeat detection to the Python intrusion detection, at the same time, the PLC data processing end carries out heartbeat detection to the C# data transmission interface, which can timely find the abnormality of the intrusion detection system, and effectively improves the safety of the on-site staff and production vehicles.

[0066] The image of the area with molten iron light in the steel plant is enhanced by the molten iron light image enhancement method in the steel plant, and in the enhancement process, the selection mode of more than one area on the image in the steel plant in step S4 of the molten iron light image enhancement method in the steel plant is randomly selected, and the randomness of random selection makes the enhanced image complex and varied, and after all the enhanced images are used as the training set of the YOLOv5 model, the training data of the YOLOv5 model is greatly expanded, and the YOLOv5 model is provided with rich and diverse training data, and under the training of rich and diverse training data, the accuracy of target detection of the YOLOv5 model is greatly improved, and the reliability of intrusion detection is improved, avoiding the situation that other personnel or vehicles are not recognized during intrusion detection due to less training set samples and low detection accuracy, and greatly improving the production safety.

[0067] The present application is suitable for ordinary skilled persons in the technical field to which the present application belongs, without departing from the concept of the present application, and a number of simple deductions or substitutions can be made, which should be considered as belonging to the protection scope of the present application.

Claims

1. A method for enhancing the illumination image of molten iron in a steel plant, characterized in that: Includes the following steps: S1. Take a picture with the light of molten iron, set the picture with the light of molten iron as picture A, and crop out the part of picture A that is affected by the light to obtain the picture of molten iron light; S2. Based on the generator network composed of deconvolutional network layers, the molten iron illumination image obtained in step S1 is used as the target of the generator network, and the generator network is optimized; the steps for optimizing the generator network include: A random vector is input into the generator network to be optimized to generate a process diagram. After calculating the loss value between the process diagram and the molten iron illumination diagram, the generator network to be optimized is updated and optimized through an optimization algorithm to obtain the optimized generator network. The above optimization steps are repeated more than once; when the optimization step is executed for the first time, the initial generator network before optimization is used as the model to be optimized; in each subsequent optimization step, the optimized generator network obtained in the previous optimization step is used as the model to be optimized. In step S2, the formula for calculating the loss value of the process diagram and the molten iron illumination diagram is as follows: ,in, The loss values ​​are those between the process diagram and the molten iron illumination diagram. This is a picture of molten iron illuminated. To generate the network, The random vector input to the generator network to be optimized. The batch size of the input random vectors. It is a positive integer; S3. Use the final optimized generator network obtained in step S2 to generate a feature map of molten iron illumination; S4. Obtain an image inside the steel plant, select one or more regions on the image inside the steel plant, and attach the molten iron illumination feature map generated in step S3 to the corresponding region of the image inside the steel plant to enhance the image inside the steel plant. In step S4, the molten iron illumination feature map generated in step S3 is appended to the corresponding region of the steel plant image using either a linear blending algorithm or a Poisson fusion algorithm. When appending the molten iron illumination feature map generated in step S3 to the corresponding region of the steel plant image using a linear blending algorithm, the formula for the linear blending algorithm is: ,in, For the enhanced image, Images of the interior of a steel plant. This is a diagram showing the illumination characteristics of molten iron. To add weights, The value ranges from 0 to 1.

2. The method for enhancing the illumination image of molten iron in a steel plant according to claim 1, characterized in that: When the size of the cut molten iron illumination image is inconsistent with the two-dimensional size of the process diagram, before calculating the loss value between the process diagram and the molten iron illumination image, the cut molten iron illumination image is scaled using bilinear interpolation to make the size of the molten iron illumination image consistent with the two-dimensional size of the process diagram output by the generating network.

3. The method for enhancing the illumination image of molten iron in a steel plant according to claim 1, characterized in that: In step S2, the optimization algorithm is any one of the following: stochastic gradient descent, Nesterov gradient acceleration, or Adam adaptive moment estimation.

4. The method for enhancing the illumination image of molten iron in a steel plant according to claim 1, characterized in that: In step S3, when generating the molten iron illumination feature map using the final optimized generator network from step S2, the input of the final optimized generator network is a random vector.

5. An intrusion detection method for detecting intrusions in areas of a steel plant exposed to molten iron light, characterized in that: Includes the following steps: J1. Batch acquire images of areas within a steel plant that are illuminated by molten iron, and use these images as images B; enhance all images B using the molten iron illumination image enhancement method described in any one of claims 1-4, wherein in step S4 of the molten iron illumination image enhancement method, one or more areas are randomly selected from the images within the steel plant; J2. Based on the YOLOv5 model, all the enhanced images B are used as the training set and input into the YOLOv5 model to train the YOLOv5 model; J3. Real-time acquisition of detection images of the area to be detected within the steel plant. The real-time acquisition of detection images of the area to be detected within the steel plant is used as image C. The YOLOv5 model trained in step J2 is used to perform target detection on image C. If a target is detected, the location of the target in image C is obtained, and it is determined whether the target is located within the defined danger zone, so as to realize intrusion detection of the area to be detected within the steel plant.

6. The intrusion detection method according to claim 5, characterized in that: In step J2, the specific steps for training the YOLOv5 model include: iteratively inputting the training set data into the model, calculating the loss function of the YOLOv5 model, and continuously adjusting the weights of the YOLOv5 model through backpropagation.

7. The intrusion detection method according to claim 6, characterized in that: In step J3, the ray method, parity rule, or non-zero rule are used to determine whether the target is located within the designated danger zone.

8. The intrusion detection method according to claim 6, characterized in that: The intrusion detection method is based on an intrusion detection system, which includes an intrusion detection camera, a Python intrusion detection system, a C# data transmission interface, a PLC data processing terminal, and a WinCC display unit. The C# data transmission interface and the Python intrusion detection system start automatically after power-on. After the intrusion detection system is started, the Python intrusion detection system periodically sends a normal operation signal to the C# data transmission interface. When the C# data transmission interface fails to receive the normal operation signal sent by the Python intrusion detection system in a timely manner, it is considered that the Python intrusion detection system is abnormal. The C# data transmission interface sends the abnormal signal of the Python intrusion detection system to the PLC data processing end, and the PLC data processing end sends it to the WinCC display unit for display. The C# data transmission interface periodically sends a normal operation signal to the PLC data processing terminal. If the PLC data processing terminal does not receive the normal operation signal from the C# data transmission interface in a timely manner, it considers the C# data transmission interface to be abnormal and sends the abnormal signal to the WinCC display unit for display.

Images