A stockyard belt deviation detection method and system based on visual recognition

Abstract

The application discloses a kind of based on visual identification's material yard belt deviation detection method and system, belong to raw material yard transport technical field, including: real-time acquisition belt running video, current frame image is input to target detection network and carries out image processing;The image data set of belt normal operation and deviation is input to network training, obtain the target detection network based on neural network as the detector of belt carrier roller, the current frame image of real-time acquisition belt running video is carried out target detection;When the current frame image cannot detect target carrier roller and alarm deviation, real-time acquisition picture in the belt edge position and calculate compared with the offset of belt normal operation;The current frame picture detection result of real-time belt running video by target detection network and after image algorithm processing offset, notify operator in time processing by the mode of alarm.The application realizes the automation and unmanned function of real-time dynamic detection of belt transport state.

Description

Technical Field

[0001] This invention belongs to the field of raw material yard transportation technology, and more specifically, relates to a method and system for detecting belt misalignment in a material yard based on visual recognition. Background Technology

[0002] In China, coal is a vital energy source, and the coal mining industry plays a significant role in promoting and driving national economic development. In the mining and coal powder industry, belt conveyors are crucial transportation equipment, responsible for conveying materials in raw material yards. However, in some actual production environments, long-term operation can lead to belt aging and loosening, uneven stress, coal sticking to rollers, or belt vibration causing radial runout of the rollers, resulting in uneven stress along the belt width and belt misalignment. When severe belt misalignment occurs during material transport, all the material on the belt can spill out, causing significant impact and losses on production. Emergency shutdown and repositioning repairs are necessary, along with handling the spilled material, wasting considerable manpower and resources. Therefore, ensuring the normal operation of the belt and preventing serious consequences caused by belt misalignment is of paramount importance.

[0003] Common belt conveyor self-balancing devices, such as belt alignment devices, play a role in preventing belt misalignment to some extent, but they cannot completely prevent or eliminate belt misalignment. Operators still need to monitor the belt operation in real time using a remote monitoring system and take emergency measures to stop severely misaligned belts. However, relying on operators to monitor the scene in real time for extended periods is inefficient due to the numerous and complex real-time images transmitted from various locations. Fatigue from prolonged monitoring can also lead to missed detections. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention proposes a method and system for detecting belt misalignment in material yards based on visual recognition. This method can solve the problem of manually monitoring serious belt misalignment during the material conveying process in material yards. By using target detection and traditional image recognition methods, the misalignment can be monitored and alarmed in real time, which can improve detection efficiency and save manpower.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for detecting belt misalignment in a material yard based on visual recognition is provided, comprising:

[0006] The current frame image in the belt running video stream is acquired in real time and input into the target detection network model. The belt idler roller is used as the detection target to determine whether the belt is running off-track.

[0007] When the result obtained by the target detection network model is belt misalignment, the current frame image in the real-time acquired belt running video stream is processed to calculate the belt offset during the belt running process in real time.

[0008] Output alarm information, including belt misalignment and belt offset.

[0009] In some optional implementations, the object detection network model is trained as follows:

[0010] Video images of belt conveyors during normal operation and during belt misalignment are collected to form historical data of belt misalignment video images;

[0011] Historical data is used as samples and input into the target detection network model for training, resulting in a deep convolutional neural network model based on neural networks, which serves as the target detection network model for belt misalignment.

[0012] In some optional implementations, the collection of video images of the belt conveyor during normal operation and during belt misalignment to form historical data of belt misalignment video images includes:

[0013] Positioning and calibrating the idler rollers under the belt during normal operation and when belt deviation occurs, to determine the coordinate position of the idler rollers;

[0014] The acquired video images are labeled with the corresponding idler coordinates and target category labels to obtain labeled historical data. The target category labels indicate whether the belt is misaligned.

[0015] The labeled historical data is divided into a training set and a test set. The training set is used to train the object detection network model and obtain network parameters, while the test set is used to test and verify the detection performance of the object detection network model.

[0016] In some optional implementations, the step of using historical data as samples to train a target detection network model to obtain a deep convolutional neural network model based on a neural network as the target detection network model for belt misalignment includes:

[0017] The training set is used as input to the object detection network model. The structure of the object detection network model is modified and the training parameters are set according to the requirements. The initial object detection network model is obtained after training.

[0018] The test set is used as input to the initial target detection network model. The trained target detection network model is then used for calculation and verification to determine whether the target category label and the idler coordinate position are consistent with the detection output of the target detection network model. If they are consistent, it indicates that the detection result of the target detection network model matches the actual result. The judgment results are statistically analyzed. When the matching accuracy of the idler is greater than a set threshold, the current target detection network model is determined to be the target detection network model for belt misalignment. When the matching accuracy of the idler is less than or equal to the set threshold, the current initial target detection network model does not meet the condition, and the training operation continues until the matching accuracy of the idler obtained from the test set meets the condition.

[0019] In some optional implementations, the processing of the current frame image in the real-time acquired belt operation video stream to calculate the belt offset during belt operation includes:

[0020] The current frame image is binarized to obtain a grayscale image. A region of interest (ROI) is selected for the target detection area in the grayscale image. Within the ROI, the belt edge contour features are extracted. The extracted belt edge contour is fitted with a straight line to obtain the position information of the points on the belt edge.

[0021] By annotating the belt scene on-site, the relationship between the belt size and the image pixel ratio in the actual scene is obtained, and thus the actual position of the belt edge is obtained;

[0022] The belt offset is calculated by comparing the position information of the points on the belt edge with the actual position of the belt edge during normal operation.

[0023] In some optional implementations, the target detection network model consists of multiple convolutional layers connected by multiple residual skip layers. It also employs multi-scale feature, cross-scale feature fusion, and upsampling methods to enhance the accuracy of small target detection. Specifically, the target confidence loss and target category loss use binary cross-entropy loss, and the predicted object category uses logistic regression to score the targetness of the predicted location, determining the probability value of the region as a target. The output includes the detection of the idler under the belt and the probability of the idler's position and category.

[0024] According to another aspect of the present invention, a vision-based belt misalignment detection system for material yards is provided, comprising:

[0025] The belt misalignment image acquisition module is used to acquire the current frame image in the belt running video stream in real time. The acquired current frame image in the belt running video stream is input into the target detection network model, with the belt under roller as the detection target, to determine whether the belt is misaligned.

[0026] The image processing module is used to process the current frame image in the real-time acquired video stream of the belt running when the result obtained by the target detection network model is belt misalignment, and to calculate the belt offset during the belt running process in real time.

[0027] An alarm module is used to output alarm information, including belt misalignment and belt offset.

[0028] In some optional implementations, the system further includes: an object detection network module;

[0029] The image processing module is also used to collect video images of the belt conveyor during normal operation and during belt misalignment, forming historical data of belt misalignment video images;

[0030] The target detection network module is used to input historical data as samples into the target detection network model for training, and obtain a deep convolutional neural network model based on neural networks as the target detection network model for belt misalignment.

[0031] In some optional implementations, the image processing module is used to locate and calibrate the idler rollers below the belt during normal operation and when belt deviation occurs, and determine the coordinate position of the idler rollers; to affix the corresponding idler roller coordinate position and target category label to each image in the acquired video images to obtain labeled historical data, wherein the target category label reflects whether the belt is deviating; the labeled historical data is divided into a training set and a test set, the training set is used to train the target detection network model and obtain network parameters, and the test set is used to test and verify the detection effect of the target detection network model.

[0032] In some optional implementations, the target detection network module is used to take the training set as input to the target detection network model, modify the structure of the target detection network model and set training parameters as needed, and obtain an initial target detection network model through training; take the test set as input to the initial target detection network model, and perform calculation verification through the trained target detection network model to determine whether the target category label and the idler coordinate position are consistent with the detection output result of the target detection network model. If they are consistent, it indicates that the detection result of the target detection network model matches the actual result, and the judgment result is statistically analyzed. When the matching accuracy of the idler is greater than a set threshold, the current target detection network model is determined to be the target detection network model for belt misalignment. When the matching accuracy of the idler is less than or equal to the set threshold, the current initial target detection network model does not meet the condition, and the training operation continues until the matching accuracy of the idler obtained from the test set meets the condition.

[0033] In some optional implementations, the image processing module is used to binarize the current frame image to obtain a grayscale image, select a Region of Interest (ROI) for the target detection area in the grayscale image, extract belt edge contour features within the ROI, perform straight line fitting on the extracted belt edge contour to obtain the position information of points on the belt edge; by annotating the belt scene on-site, obtain the ratio of belt size to image pixels in the actual scene, and thus obtain the actual position of the belt edge; compare the position information of points on the belt edge with the actual position of the belt edge during normal operation to calculate the belt offset.

[0034] In some optional implementations, the target detection network model consists of multiple convolutional layers connected by multiple residual skip layers. It also employs multi-scale feature, cross-scale feature fusion, and upsampling methods to enhance the accuracy of small target detection. Specifically, the target confidence loss and target category loss use binary cross-entropy loss, and the predicted object category uses logistic regression to score the targetness of the predicted location, determining the probability value of the region as a target. The output includes the detection of the idler under the belt and the probability of the idler's position and category.

[0035] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0036] This invention involves real-time acquisition of belt conveyor operation video, inputting the current frame image into a target detection network for image processing. By training the network with image datasets of both normal belt operation and belt misalignment, a neural network-based target detection network is obtained and acts as a detector for belt idlers. This network detects targets in the current frame of the real-time acquired belt conveyor video. If no target idler is detected in the current frame, an alarm for belt misalignment is triggered. The network then acquires the belt edge position in real-time and calculates the offset compared to normal belt operation. The detection results of the target detection network for the current frame of the real-time belt conveyor video, along with the offset after image algorithm processing, are used to notify operators via an alarm for timely handling. This invention achieves automated and unmanned real-time dynamic detection of belt conveyor status, improving detection efficiency and saving manpower. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating a vision-based belt misalignment detection method for a material yard, as provided in an embodiment of the present invention.

[0038] Figure 2 This is a schematic diagram of a vision recognition-based conveyor belt misalignment detection system provided in an embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0040] This invention designs a belt misalignment detection method and system based on computer vision recognition technology. It uses a target detection method with a high recognition rate, combined with traditional image recognition technology to provide feedback on the degree of misalignment, which saves a lot of manpower and improves work efficiency and detection efficiency.

[0041] Example 1

[0042] like Figure 1 The diagram shown is a flowchart illustrating a vision-based belt misalignment detection method for a material yard, as provided in an embodiment of the present invention, including:

[0043] The current frame image in the belt running video stream is acquired in real time and input into the target detection network model. The belt idler roller is used as the detection target to determine whether the belt is running off-track.

[0044] When the result obtained by the target detection network model is belt misalignment, the current frame image in the real-time acquired belt running video stream is processed to calculate the belt offset during the belt running process in real time.

[0045] Output alarm information, including belt misalignment and belt offset.

[0046] In this embodiment, the target detection network model is trained as follows:

[0047] Video images of belt conveyors during normal operation and during belt misalignment are collected to form historical data of belt misalignment video images;

[0048] Historical data is used as samples and input into the target detection network model for training, resulting in a deep convolutional neural network model based on neural networks, which serves as the target detection network model for belt misalignment.

[0049] In this embodiment, video images of the belt conveyor during normal operation and during belt misalignment are collected to form historical data of belt misalignment video images, including:

[0050] Positioning and calibrating the idler rollers under the belt during normal operation and when belt deviation occurs, to determine the coordinate position of the idler rollers;

[0051] The corresponding idler coordinate position and target category label are affixed to each image in the acquired video image to obtain labeled historical data, where the target category reflects whether the belt is running off-track;

[0052] The labeled historical data is divided into a training set and a test set. The training set is used to train the object detection network model and obtain network parameters, while the test set is used to test and verify the detection performance of the object detection network model.

[0053] In this embodiment, historical data is used as samples and input into the target detection network model for training, resulting in a deep convolutional neural network model based on a neural network, which serves as the target detection network model for belt misalignment. This model includes:

[0054] The training set is used as input to the object detection network model. The structure of the object detection network model is modified and the training parameters are set according to the requirements. The initial object detection network model is obtained after training.

[0055] The test set is used as input to the initial target detection network model. The trained target detection network model is then used for calculation and verification to determine whether the target category label and the idler coordinate position are consistent with the detection output of the target detection network model. If they are consistent, it indicates that the detection result of the target detection network model matches the actual result. The judgment results are statistically analyzed. When the matching accuracy of the idler is greater than a set threshold, the current target detection network model is determined to be the target detection network model for belt misalignment. When the matching accuracy of the idler is less than or equal to the set threshold, the current initial target detection network model does not meet the condition, and the training operation continues until the matching accuracy of the idler obtained from the test set meets the condition.

[0056] In this embodiment, the current frame image in the real-time acquired video stream of the belt conveyor operation is processed to calculate the belt offset during the belt operation process in real time, including:

[0057] The current frame image is binarized to obtain a grayscale image. A region of interest (ROI) is selected for the target detection area in the grayscale image. Within the ROI, the belt edge contour features are extracted. The extracted belt edge contour is fitted with a straight line to obtain the position information of the points on the belt edge.

[0058] By annotating the belt scene on-site, the relationship between the belt size and the image pixel ratio in the actual scene is obtained, and thus the actual position of the belt edge is obtained;

[0059] The belt offset is calculated by comparing the position information of the points on the belt edge with the actual position of the belt edge during normal operation.

[0060] In this embodiment, the target detection network model consists of multiple convolutional layers and uses multiple residual skip-layer connections. It also employs multi-scale feature, cross-scale feature fusion, and upsampling methods to enhance the accuracy of small target detection. Specifically, the target confidence loss and target category loss use binary cross-entropy loss, and the predicted object category uses logistic regression to score the targetness of the predicted location, determining the probability value of the region as a target, and outputting the detection of the idler roller under the belt and the probability of the roller's position and category.

[0061] Example 2

[0062] This embodiment discloses a vision recognition-based conveyor belt misalignment detection system for material yards, such as... Figure 2 As shown, it includes: a belt misalignment image acquisition module, a target detection network module, an image processing module, and an alarm module; wherein:

[0063] The belt misalignment image acquisition module is used to collect video images of the belt conveyor during normal operation and during belt misalignment due to faults, forming historical data of belt misalignment video images; it is also used to acquire the current frame image in the belt running video stream in real time, and input the current frame image in the real-time acquired belt running video stream into the target detection network model for calculation;

[0064] In this embodiment, the belt misalignment image acquisition module can be a camera positioned directly above the belt. The specific method for the belt misalignment image acquisition module to collect images of the belt in operation and form a historical dataset is as follows: the belt misalignment image acquisition module locates and calibrates the idler rollers below the belt during normal operation and when misalignment occurs; each image in the acquired dataset is labeled with the corresponding idler roller coordinate position information and target category label; the labeled dataset is divided into two categories: one is a training set, used to train the target detection network model and obtain network parameters; the other is a test set, used to test and verify the detection effect of the target detection network model.

[0065] The target detection network module uses image data of the belt during normal operation and belt deviation as samples to train the target detection network model. The resulting deep convolutional neural network model serves as a detector for belt deviation. It detects the current frame image in the real-time acquired belt operation video, using the belt idler roller as the detection target. When the belt deviates, it covers one side of the idler roller, while the other side of the idler roller is largely exposed. Based on the detection results, it achieves the purpose of real-time monitoring of whether the belt has deviated.

[0066] In this embodiment, the training method for the object detection network model includes: using a portion of the training set in the image dataset as input to the object detection network model; modifying the structure of the object detection network model and setting training parameters according to requirements; and obtaining an initial object detection network model through training. Then, using a test set in the image dataset as input to the initial object detection network model, the trained object detection model network is used for calculation and verification to determine whether the label and location information are consistent with the detection output results and to statistically analyze the results. When the matching accuracy of the idler roller is greater than a set threshold, the current object detection network model is determined to be the object detection network model for belt misalignment. The training method for the object detection network model further includes: when the matching accuracy is less than or equal to the set threshold, the current initial object detection network model does not meet the conditions, and the above training steps need to be continued until the matching degree of the test set meets the requirements.

[0067] In this embodiment, the target detection network model consists of multiple convolutional layers and extensively uses residual skip connections. Specifically, the target detection network model contains 53 convolutional layers, with an input image size of 256*256*3. It uses convolutions with a stride of 2 for downsampling and employs multi-scale feature, cross-scale feature fusion, and upsampling methods to enhance the accuracy of small target detection.

[0068] In this embodiment, the target confidence loss and target category loss adopt the binary cross-entropy loss. The predicted object category uses logistic regression to score the targetness of the predicted location and determine the probability value of the area being the target, thus completing the detection and output of the position and probability of the idler roller under the belt.

[0069] The image processing module processes the current frame image in the real-time acquired video of the belt running and calculates the offset of the belt during the running process in real time.

[0070] In this embodiment, the image processing module processes the current frame image in the real-time acquired belt running video. The processing method includes: binarizing the current frame image to obtain a grayscale image, selecting a region of interest (ROI) for the target detection area, extracting belt edge contour features within the ROI, fitting the extracted belt edge contour with a straight line, and obtaining the position information of points on the belt edge.

[0071] In this embodiment, the target belt scene is annotated on-site to obtain the ratio of the target size in the actual scene to the image pixels in the belt deviation image acquisition module. Through mathematical calculation, the actual position of the belt edge is returned. When the result obtained by the target detection network model is deviation, the difference between the real-time belt edge position and the position during normal operation is compared to calculate the offset.

[0072] The alarm module obtains the current frame image processing result of the target detection network model on the real-time acquired belt running video and the belt offset output by the image processing module. If the detection result of the processed image frame is belt deviation, the alarm module will notify the relevant person in charge in real time and provide a numerical analysis of the degree of deviation to ensure that the relevant personnel can deal with it in a timely manner.

[0073] In this embodiment, the alarm module can be a buzzer. When the target detection network model detects that the current frame image in the real-time acquired video of the belt running is off-track, it outputs the belt offset and alarms in real time through the buzzer, notifying the relevant person in charge to confirm the degree of deviation and handle it in a timely manner.

[0074] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0075] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting belt misalignment in a material yard based on visual recognition, characterized in that, include: The system acquires the current frame image from the real-time video stream of the belt operation and inputs it into a target detection network model. The target detection network model consists of multiple convolutional layers and uses multiple residual skip connections. It employs multi-scale feature fusion, cross-scale feature fusion, and upsampling methods. The target confidence loss and target category loss use binary cross-entropy loss. The predicted object category uses logistic regression to score the targetness of the predicted location. The idler roller under the belt is used as the detection target, and the probability value of the region being the target is determined. The target detection network model outputs the detection of the idler roller under the belt, as well as the roller position and category probability, to determine whether the belt is off-track. When the target detection network model calculates that the belt is misaligned, the current frame image in the real-time acquired belt operation video stream is processed. The current frame image is binarized to obtain a grayscale image. A Region of Interest (ROI) is selected for the target detection area in the grayscale image. Within the ROI, the belt edge contour features are extracted. The extracted belt edge contour is fitted with a straight line to obtain the position information of the points on the belt edge. By annotating the belt scene on-site, the ratio of belt size to image pixels in the actual scene is obtained, thereby obtaining the actual position of the belt edge. The position information of the points on the belt edge is compared with the actual position of the belt edge during normal operation, and the belt offset during the belt operation is calculated in real time. Output alarm information, including belt misalignment and belt offset; The training method for the object detection network model is as follows: Video images of belt conveyors during normal operation and during belt misalignment are collected to form historical data of belt misalignment video images; Historical data is used as samples and input into the target detection network model for training, resulting in a deep convolutional neural network model based on neural networks, which serves as the target detection network model for belt misalignment. The collection of video images of the belt conveyor during normal operation and during belt misalignment, forming historical data of belt misalignment video images, includes: Positioning and calibrating the idler rollers under the belt during normal operation and when belt deviation occurs, to determine the coordinate position of the idler rollers; The acquired video images are labeled with the corresponding idler coordinates and target category labels to obtain labeled historical data. The target category labels indicate whether the belt is misaligned. The labeled historical data is divided into a training set and a test set. The training set is used to train the object detection network model and obtain the network parameters, while the test set is used to test and verify the detection performance of the object detection network model. The process of using historical data as samples and inputting it into the target detection network model for training, resulting in a deep convolutional neural network model based on a neural network as the target detection network model for belt misalignment, includes: The training set is used as input to the object detection network model. The structure of the object detection network model is modified and the training parameters are set according to the requirements. The initial object detection network model is obtained after training. The test set is used as input to the initial target detection network model. The trained target detection network model is then used for calculation and verification to determine whether the target category label and the idler coordinate position are consistent with the detection output of the target detection network model. If they are consistent, it indicates that the detection result of the target detection network model matches the actual result. The judgment results are statistically analyzed. When the matching accuracy of the idler is greater than a set threshold, the current target detection network model is determined to be the target detection network model for belt misalignment. When the matching accuracy of the idler is less than or equal to the set threshold, the current initial target detection network model does not meet the condition, and the training operation continues until the matching accuracy of the idler obtained from the test set meets the condition.

2. A vision-based conveyor belt misalignment detection system for material yards, characterized in that, include: The belt misalignment image acquisition module is used to acquire the current frame image in the belt running video stream in real time. The acquired current frame image in the belt running video stream is input into the target detection network model. The target detection network model consists of multiple convolutional layers and uses multiple residual skip-layer connections. It adopts multi-scale feature, cross-scale feature fusion and upsampling methods. The target confidence loss and target category loss adopt binary cross-entropy loss. The predicted object category uses logistic regression to score the targetness of the predicted position. The idler under the belt is used as the detection target, and the probability value of the region is judged as the target. The target detection network model outputs the detection of the idler under the belt and the idler position and category probability to determine whether the belt is misaligned. The image processing module is used to process the current frame image in the real-time acquired belt running video stream when the result obtained by the target detection network model is belt deviation. The current frame image is binarized to obtain a grayscale image. A region of interest (ROI) is selected for the target detection area in the grayscale image. Within the ROI, belt edge contour features are extracted. The extracted belt edge contour is fitted with a straight line to obtain the position information of the points on the belt edge. By annotating the belt scene on site, the ratio of belt size to image pixels in the actual scene is obtained, thereby obtaining the actual position of the belt edge. The position information of the points on the belt edge is compared with the actual position of the belt edge during normal operation to calculate the belt offset in real time during belt operation. An alarm module is used to output alarm information, including belt misalignment and belt offset. The system also includes: a target detection network module; The image processing module is also used to collect video images of the belt conveyor during normal operation and during belt misalignment, forming historical data of belt misalignment video images; The target detection network module is used to input historical data as samples into the target detection network model for training, and obtain a deep convolutional neural network model based on neural network as the target detection network model for belt deviation. The image processing module is used to locate and calibrate the idler rollers under the belt during normal operation and when belt deviation occurs, and to determine the coordinate position of the idler rollers; it affixes the corresponding idler roller coordinate position and target category label to each image in the acquired video image to obtain labeled historical data, wherein the target category label reflects whether the belt is deviating; the labeled historical data is divided into a training set and a test set, the training set is used to train the target detection network model and obtain network parameters, and the test set is used to test and verify the detection effect of the target detection network model; The target detection network module is used to take the training set as input to the target detection network model, modify the structure of the target detection network model and set training parameters according to requirements, and obtain an initial target detection network model through training. The test set is used as input to the initial target detection network model, and the trained target detection network model is used for calculation and verification to determine whether the target category label and the idler coordinate position are consistent with the detection output of the target detection network model. If they are consistent, it indicates that the detection result of the target detection network model matches the actual result, and the judgment results are statistically analyzed. When the matching accuracy of the idler is greater than a set threshold, the current target detection network model is determined to be the target detection network model for belt misalignment. When the matching accuracy of the idler is less than or equal to the set threshold, the current initial target detection network model does not meet the condition, and the training operation continues until the matching accuracy of the idler obtained from the test set meets the condition.

Images