Plate initial welding position segmentation method and system based on BoT3-yolov8

Abstract

The application discloses a medium plate initial welding position segmentation method and system based on BoT3-YOLOv8, which comprises the following steps: collecting welding seam image data of a medium plate; pre-processing the welding seam image data to obtain processed image data; training a network model of a welding position segmentation model by using the processed image data to obtain a trained welding position segmentation model; inputting a to-be-tested welding seam image into the trained welding position segmentation model to output initial welding position feature information; and performing mask processing on the initial welding position feature information to obtain an initial welding position mask binary image. The application has higher detection precision, stronger anti-interference capability, further improves the welding efficiency of the medium plate, and has a wide application range.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of welding segmentation, in particular to a middle-thick plate initial welding position segmentation method and system based on BoT3-YOLOv8. BACKGROUND

[0002] Welding seam initial position information acquisition and welding robot visual guidance are one of the key technologies to realize automatic welding. However, the middle-thick plate has the characteristics of large space volume, irregular welding seam and discrete welding seam distribution, which makes it difficult for the welding robot to realize visual guidance of the initial space position of the middle-thick plate welding seam alone. Therefore, it is necessary to extract the features of the initial welding position of the middle-thick plate.

[0003] At present, the main directions are active vision and passive vision for feature extraction of welding seam initial position. In the active vision research direction, line structured light sensors have been widely used due to their good robustness and high precision, such as welding seam initial position recognition, parameter extraction, welding seam tracking, and welding pool monitoring. Liu F et al. proposed a laser inflection point automatic extraction algorithm based on dynamic programming to help welding robots obtain the initial position of fillet welds. Wang N et al. proposed a welding seam recognition method based on structured light in strong noise, which generates a dynamic region of interest using the NURBS-snake model to extract the welding seam center line. Li W et al. used a welding seam edge point detection operator and IPCE algorithm to identify multiple types of welding seams. Although the welding seam recognition method based on line structured light sensors has strong anti-interference ability, the information acquisition speed of such methods is slow. Line structured light sensors can only acquire one row or one column of three-dimensional information of the target object at a time, and the complete three-dimensional information of the entire measurement area can only be obtained through motion scanning. In addition, the relative uniformity between the target object and the measuring instrument must be ensured during scanning, otherwise the shape will be distorted, affecting the imaging quality. In summary, the active vision method cannot guarantee the welding efficiency of the middle-thick plate with large space volume.

[0004] In the passive vision research direction, deep learning combined with welding robot perception technology becomes the dominant, and most of the technologies gradually use reinforcement learning and self-supervised learning methods to identify welding seam features. In order to improve the efficiency of automatic welding, some researchers first use laser sensors to collect welding seam images, and then let the welding robot perform reinforcement learning and self-supervised learning from the welding seam image data to realize welding seam feature recognition and extraction. Another part of the researchers establishes a welding seam feature extraction model through the method of 3D point cloud, and then completes the path planning and posture planning of the welding robot. Although the welding seam recognition technology combining deep learning and passive vision improves the efficiency of automatic welding, the above-mentioned technologies in the passive vision research direction can only recognize a single type of welding seam, and have low applicability to complex middle-thick plate welding seams.

[0005] Therefore, there is a need for a BoT3-YOLOv8-based initial welding position segmentation method and system for medium-thick plates, which can solve the above problems. SUMMARY

[0006] Therefore, the purpose of the present application is to overcome the defects in the prior art, provide a BoT3-YOLOv8-based initial welding position segmentation method and system for medium-thick plates, which has higher detection accuracy and stronger anti-interference ability, further improves the welding efficiency of medium-thick plates, and has a wide range of applications.

[0007] The BoT3-YOLOv8-based initial welding position segmentation method for medium-thick plates of the present application comprises the following steps:

[0008] Collecting welding image data of the medium-thick plate;

[0009] Preprocessing the welding image data to obtain processed image data;

[0010] Training the welding position segmentation model using the processed image data to obtain a trained welding position segmentation model;

[0011] Inputting the welding image to be tested into the trained welding position segmentation model to output initial welding position feature information;

[0012] Mask processing the initial welding position feature information to obtain an initial welding position mask binary image.

[0013] Further, the welding position segmentation model comprises an improved YOLOv8 model;

[0014] The improved YOLOv8 model is based on the YOLOv8 model, fuses the Bottleneck Transformer into the bottleneck layer of C2f to obtain the C2f_Bottleneck_BoT3 module, and applies the C2f_Bottleneck_BoT3 module to the neck network of YOLOv8.

[0015] Further, the C2f_Bottleneck_BoT3 module adopts a multi-head self-attention mechanism.

[0016] Further, the improved YOLOv8 model uses a dilated convolution module as a spatial pyramid pooling structure connecting the backbone network and the neck network of the YOLOv8 model.

[0017] Further, for a given size of image input, the dilated convolution module uses different sampling rates of dilated convolution to sample the input image, then splices the output results of each dilated convolution at the beginning and end, expands the channel number of the image, and finally reduces the channel number of the image through 1x1 convolution to output the image.

[0018] Further, the initial welding position feature information is subjected to mask processing by adopting an HSV color space region segmentation method, specifically including:

[0019] The color value interval of the weld mask is determined by the H component histogram;

[0020] The value range of the H component is set to 0-25, and then the S component and the V component are fine-tuned, and the mask binary image of the initial welding position feature of the medium plate is generated according to the color limit of the HSV model;

[0021] Among them, for the noise in the generated mask binary image, an erosion and expansion processing method is adopted to denoise the mask binary image.

[0022] A medium plate initial welding position segmentation system based on BoT3-YOLOv8 includes a collection unit, a preprocessing unit, a model training unit, a welding position segmentation unit, and a welding mask unit.

[0023] The collection unit is configured to collect welding seam image data of a medium plate.

[0024] The preprocessing unit is configured to preprocess the welding seam image data to obtain processed image data.

[0025] The model training unit is configured to train a welding position segmentation model using the processed image data to obtain a trained welding position segmentation model.

[0026] The welding position segmentation unit is configured to input a welding seam image to be tested into the trained welding position segmentation model to output initial welding position feature information.

[0027] The welding mask unit is configured to perform mask processing on the initial welding position feature information to obtain an initial welding mask binary image.

[0028] Further, the welding position segmentation model includes an improved YOLOv8 model.

[0029] The improved YOLOv8 model is based on the YOLOv8 model, fuses the Bottleneck Transformer into the bottleneck layer of C2f, obtains the C2f_Bottleneck_BoT3 module, and applies the C2f_Bottleneck_BoT3 module to the neck network of YOLOv8.

[0030] Further, the C2f_Bottleneck_BoT3 module adopts a multi-head self-attention mechanism.

[0031] Further, the improved YOLOv8 model uses a cavity convolution module as a spatial pyramid pooling structure connecting the YOLOv8 model backbone network and the neck network.

[0032] The beneficial effects of the present application are: the disclosed middle plate initial welding position segmentation method and system based on BoT3-YOLOv8 combine deep learning algorithm and image segmentation technology to establish a middle plate initial welding position instance segmentation model; by designing a new gradient shunt module C2f_Bottleneck_BoT3 and using cavity convolution as the spatial pyramid pooling structure of the segmentation model, the feature expression ability and target multi-scale information acquisition ability of the segmentation model are enhanced, the HSV color space region segmentation method is used for post-processing of the prediction result of the model, and a weld feature binary image is obtained. The detection precision of the present application is high, the anti-interference ability is strong, the single weld image processing speed is fast, and the application scope is wide. BRIEF DESCRIPTION OF DRAWINGS

[0033] The present application will be further described below in conjunction with the drawings and examples:

[0034] Figure 1 The initial welding position segmentation method of the present application is shown in the schematic diagram;

[0035] Figure 2 The improved YOLOv8 model network structure of the present application is shown in the schematic diagram;

[0036] Figure 3 The C2f_Bottleneck_BoT3 module structure of the present application is shown in the schematic diagram;

[0037] Figure 4 The SA and MHSA structure of the present application is shown in the schematic diagram;

[0038] Figure 5 The cavity convolution module ASPP structure of the present application is shown in the schematic diagram;

[0039] Figure 6 The H component histogram of the present application is shown in the schematic diagram;

[0040] Figure 7 The mask image of the present application is shown in the schematic diagram;

[0041] Figure 8 The mask binary image of the present application is shown in the schematic diagram;

[0042] Figure 9 The mask binary image erosion denoising effect of the present application is shown in the schematic diagram;

[0043] Figure 10 The mask binary image inflation denoising effect of the present application is shown in the schematic diagram. DETAILED DESCRIPTION

[0044] The application is further described in the following description in connection with the attached drawings, in which:

[0045] The BoT3-YOLOv8-based plate initial welding position segmentation method of the application comprises the following steps:

[0046] Collecting welding seam image data of the plate;

[0047] Pretreating the welding seam image data to obtain processed image data;

[0048] Training a welding position segmentation model using the processed image data to obtain a trained welding position segmentation model;

[0049] Inputting the welding seam image to be tested into the trained welding position segmentation model to output initial welding position feature information;

[0050] Mask processing the initial welding position feature information to obtain an initial welding position mask binary image.

[0051] The application adopts a YOLOv8 algorithm to build a plate initial welding position instance segmentation basic model, fuses a Bottleneck Transformer into a C2f to obtain a C2f_Bottleneck_BoT3 module, and uses the module to filter redundant information in the image and improve the feature expression capability of the model. The Bottleneck Transformer is a neural network structure based on an attention mechanism. Then, an empty convolution is used as a spatial pyramid pooling structure to connect a backbone network and a neck network to obtain target multi-scale information. Compared with other methods, the application has higher detection accuracy and stronger anti-interference capability, and provides technical support for engineering practical application. Unlike existing welding position recognition technology which only finds a welding position, the welding position segmentation technology of the application can distinguish the welding position from adjacent areas, and can highlight the target welding position.

[0052] In the model training process, a model evaluation index can be established for error analysis between training results and verification results, which is beneficial to obtain a trained weight file and further obtain a trained segmentation model. In addition, the initial welding position original image and the mask binary image are compared to correct the weight of the network model structure and further optimize the segmentation model.

[0053] In this embodiment, a depth camera installed near a welding robot welding gun can be used to collect welding seam image data; then, the collected image data is subjected to normalization and data enhancement and other pretreatment operations; the data enhancement adopts existing image data enhancement processing technology, which will not be described here.

[0054] In this embodiment, in order to enhance the performance of the segmentation model, two optimization strategies are adopted to improve the segmentation model based on the YOLOv8 model. The specific optimization strategies are as follows:

[0055] (1) A new module is obtained by fusing Bottleneck Transformer into the bottleneck layer of C2f, and it is named C2f_Bottleneck_BoT3. In addition, all C2f modules in the neck network of YOLOv8 are replaced by C2f_Bottleneck_BoT3 modules;

[0056] (2) The Atrous Spatial Pyramid Pooling (ASPP) module is used as the spatial pyramid pooling structure connecting the Backbone part and the Neck part. The improved YOLOv8 model is as shown in Figure 2 .

[0057] Among them, the YOLOv8 network is composed of four parts: backbone network, neck, head, and prediction layer. The backbone network of YOLOv8 still continues the design idea of CSPDarknet gradient shunt. The advantage of CSPDarknet is to improve the learning ability of the network without losing accuracy and reduce the computational complexity of the network to improve the inference speed. The backbone network contains five feature clusters. The first feature cluster P1 is a 3x3 convolution, and the structures of feature clusters P2-P5 are the same, consisting of a 3x3 convolution layer and a C2f module.

[0058] The role of the neck of YOLOv8 is to input the different resolution features generated in the previous step and output the features after efficient fusion. The neck adopts the structure of PAFPN. The strategy of Path Aggregation (PA) greatly reduces the number of network levels that different levels of features need to pass through when transmitting. YOLOv8 algorithm adopts a decoupled head structure to accelerate model convergence speed and detection accuracy. The loss function of YOLOv8 includes class classification and bounding box regression. The class classification loss adopts the cross-entropy loss function (BCE), and the bounding box regression loss is composed of CIOU Loss and Distribution Focal Loss (DFL). In addition, YOLOv8 also adopts the Task-AlignedAssigner positive and negative sample dynamic allocation strategy. For the classification prediction score and regression prediction score of all pixel points, the K largest positive samples are selected after sorting the weighted scores.

[0059] The core idea of the Bottleneck Transformer model is to introduce the self-attention mechanism of the Transformer model into the bottleneck structure of the ResNet model. The Transformer Block is fused into the bottleneck layer of C2f to obtain the C2f_Bottleneck_BoT3 module, and the C2f_Bottleneck_BoT3 module is applied to the neck network of YOLOv8, thereby improving the local feature extraction capability and global information capturing capability of the initial plate welding position segmentation model. The structure of the C2f_Bottleneck_BoT3 module and its implementation process are as shown in Figure 3 .

[0060] The multi-head self-attention mechanism (MHSA) is the core module in the BoT3 structure. Compared with the self-attention mechanism (SA), MHSA adopts the design idea of divide and conquer and target information dimensionality. By fully utilizing the parallelism of the display card and the way of exchanging spatial complexity for time complexity, better model performance is obtained. The structure of MHSA is as shown in Figure 4 .

[0061] MHSA is based on SA, which generates multiple output sequences through multi-head mechanism for the input sequence, and then integrates these output sequences to obtain a new output sequence. The specific calculation process is as follows:

[0062]

[0063] Where x and z represent input and output respectively, Q is the query, K is the index, V is the content, , and the attention matrix is represented. The whole calculation process of SA is equivalent to weighted summation of the input sequence x to generate the attention matrix , and then the output sequence z is obtained. SA takes the Q, K, and V calculated before as a whole Head input to the next layer, while MHSA needs multiple groups of W q , W k , and W v to multiply with the input x to obtain multiple groups of Q, K, and V. As shown in Figure 4 , for example in the right side of the diagram, the input sequence a1 obtains three output vectors through multi-head mechanism (head=3), and then obtains the output sequence b1 through splicing and linear conversion. The calculation process is as follows:

[0064]

[0065] Where Concat represents splicing operation, and Linear represents linear conversion.

[0066] In this embodiment, the improved YOLOv8 model uses a hollow convolution module as a spatial pyramid pooling structure connecting the YOLOv8 model backbone network and the neck network.

[0067] Since the weight matrix of the full connection layer in the traditional convolutional neural network is a fixed value, the image size output by the pooling layer must be consistent, otherwise the network will not be trained. In order to solve this problem, stretching or cropping can be used to ensure that the image size output by the pooling layer is consistent, but stretching or cropping will cause image information loss and image distortion. Therefore, a pyramid pooling structure is proposed to solve the previous problem by using multi-scale image information. The pyramid pooling in the YOLOv8 model uses the SPPF structure, and the average accuracy of the pyramid pooling layer using the hollow convolution structure is higher than that of the SPPF structure in the initial welding position image instance segmentation of the medium plate. Therefore, by using the ASPP (hollow convolution module) structure as the improved pyramid pooling layer of the segmentation model, the multi-scale target information acquisition capability of the model can be enhanced. The structure of the ASPP is as shown in Figure 5 .

[0068] For a given size of image input, the ASPP module uses different sampling rates of hollow convolution to sample the input image. Then the output results of each hollow convolution are spliced at the head and tail, and the channel number of the image is expanded. Finally, the channel number is reduced by 1x1 convolution to output the image.

[0069] Through the above method, the segmentation model is improved to obtain better performance with a little increase in parameter quantity and computational complexity.

[0070] In this embodiment, in order to obtain the mask binary image of the initial welding position feature of the medium plate, the HSV color space region segmentation method is used for post-processing of the weld mask image generated by the segmentation model prediction. First, the color value interval of the weld mask is determined by the H component histogram. The distribution of the H component is as shown in Figure 6 .

[0071] As shown in Figure 6 , the color channel of the weld mask is mostly distributed between 0 and 25. Then, the value range of the H component is set to 0-25, and the S component and V component are fine-tuned to generate the mask binary image of the initial welding position feature of the medium plate according to the color limit of the HSV model (as shown in Figure 8 ). In addition, for the noise in the generated mask binary image, an erosion and expansion processing method is used to denoise the mask binary image. The results of the post-processing are as shown in Figure 9 , 10 .

[0072] The post-processed mask binary image can more clearly express the characteristics of the target object compared with the original mask image (as shown in the figure). Figure 7

[0073] The application also relates to a medium plate initial welding position segmentation system based on BoT3-YOLOv8, which corresponds to the above-mentioned medium plate initial welding position segmentation method based on BoT3-YOLOv8 and can be understood as a system for realizing the above-mentioned method, and the system comprises an acquisition unit, a preprocessing unit, a model training unit, a welding position segmentation unit and a welding position mask unit.

[0074] The acquisition unit is used for acquiring welding seam image data of the medium plate.

[0075] The preprocessing unit is used for preprocessing the welding seam image data to obtain processed image data.

[0076] The model training unit is used for training a welding position segmentation model by using the processed image data to obtain a trained welding position segmentation model.

[0077] The welding position segmentation unit is used for inputting a to-be-tested welding seam image into the trained welding position segmentation model to output initial welding position feature information.

[0078] The welding position mask unit is used for performing mask processing on the initial welding position feature information to obtain an initial welding position mask binary image.

[0079] In addition, for the medium plate initial welding position segmentation model, four indexes of precision, recall, mean average precision (mAP) and frames per second (FPS) can be used to evaluate the performance of the model.

[0080] In order to better understand the medium plate initial welding position segmentation method and system of the application, further specific description will be given as follows:

[0081] ​In order to obtain the initial position image of the weld of the medium plate workpiece, two six-degree-of-freedom Fanuc M-10iD / 8L welding robots and a depth camera are used for data acquisition. The data acquisition is mainly completed by the cooperation of the welding robot, the depth camera, the guide rail and the gantry truss. First, the gantry truss installed on the guide rail cooperates with the two mechanical arms to realize the scanning of the whole workpiece area. During the scanning process, the workpiece is photographed at different distances and angles by the depth camera. A total of 500 images are collected, and then the dataset is divided into training set, validation set and test set in the ratio of 8:1:1 according to the coco format, and the training set and the validation set are labeled by labelme. In addition, in view of the problem of image scarcity, the Mosaic data enhancement method is used to randomly crop and splice the images, so as to realize the expansion and enrichment of the dataset.

[0082] Under the win10 operating system, the hardware configuration includes an Intel(R) Core(TM) i5-12400F@2.50GHz CPU, a 12GB Nvidia GeForce RTX 3060 GPU, and a 64GB RAM. The deep learning environment used is configured as Pytorch 1.13, python3.10, cuda11.7, ultralytics8.0, and opencv-python4.6. The specific model training parameter configuration is shown in Table 1:

[0083] Table 1

[0084]

[0085] Before starting the training, the momentum factor of the model is set to 0.937, the weight decay is set to 0.0005, and the automatic mixed precision training mode is enabled. In addition, in order to prevent the model from overfitting, if the model effect still does not improve after 50 epochs of training, the training will be terminated in advance.

[0086] Under the condition of ensuring the same model training parameters, the original YOLOv8 and the BoT3-YOLOv8 of the application are compared and analyzed.

[0087] From the comparison of the loss function curve changes in the training and verification process, it can be seen that the loss function of the improved model converges faster and better, proving that the improved model has stronger learning ability for samples. By fusing the BottleneckTransformer and replacing the pyramid pooling structure, the precision, recall rate and mean average precision of the model can be significantly improved. The performance parameter comparison before and after the improvement of the model is shown in Table 2:

[0088] Table 2

[0089]

[0090] Compared with the original YOLOv8, the average precision of BoT3-YOLOv8 is improved by 5.2%, the parameter amount and the calculation amount are increased by 11.3M and 6.2GFLOPs respectively, and the FPS is reduced by 8.7. This shows that better model performance is obtained with a little increase in model parameter amount and network calculation amount, and the initial welding position detection precision of the medium plate is significantly improved.

[0091] In addition, in order to verify the superiority of the method, three kinds of weld feature extraction methods (UNet, SOLOv2, DeepLabV3+) are selected for comparison:

[0092] After 200 epochs of training, all four methods converge. Among them, the convergence speed of the method is faster, and the convergence speed of the UNet method is the slowest. In terms of model accuracy, the mAP 0.5 of the method is more than 90%. However, the mAP 0.5 of the other three methods is much lower than that of the method, the mAP 0.5 of SOLOv2 is 80%, the mAP 0.5 of DeepLabV3+ is about 75%, and the mAP 0.5 of UNet is only 60%. In summary, the method has better effect in the model training stage.

[0093] After the model training is completed, in order to verify the anti-interference ability and generalization ability of the method, four different scenes and weld types are set, and the weight file generated by model training is used for detection, and then the segmentation effect of each method is analyzed and compared.

[0094] The method has the best segmentation effect on different scenes and different types of medium plate weld images, and all target features in each weld image are accurately segmented, and the mask contour and mask shape of the method are more consistent with the initial welding position of the medium plate. Original-YOLOv8, UNet, SOLOv2 have false detection, missed detection and irregular mask shape, etc. Through the comparison of the prediction results, it is found that the DeepLabV3+ model is more sensitive to the contour of the target object, so DeepLabV3+ is easy to incorrectly identify the medium plate contour as a weld. In addition, the method adopts the HSV region segmentation method to binarize the image mask, which is convenient for the welding robot to obtain the position information of the target feature and realize the visual guidance before welding.

[0095] The method of the application is trained and predicted with the improved method and three existing methods, and then the training results of the five methods are analyzed and compared from the loss function, evaluation index and performance parameter of the model. Secondly, interference factors are added in the prediction process, and the prediction effects of the five methods on different types of weld images in different scenes are compared. The results show that:

[0096] (1) The improved model of the application has faster loss function convergence and higher average precision, and the model performance is significantly improved;

[0097] (2) Compared with other methods, the detection accuracy of the application is the highest, and the anti-interference ability and generalization ability are the best.

[0098] Finally, it should be pointed out that the above examples are only used to illustrate the technical solutions of the application and not to limit it. Although the application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the application can be modified or replaced equivalently without departing from the purpose and scope of the application, and they should be included in the scope of the claims of the application.

Claims

1. A method for plate initial welding position segmentation based on BoT3-YOLOv8, characterized in that: The method comprises the following steps: Collecting weld image data of a medium-thick plate; Preprocessing the weld image data to obtain processed image data; Training a weld position segmentation model using the processed image data to obtain a trained weld position segmentation model; the weld position segmentation model comprises an improved YOLOv8 model; The improved YOLOv8 model is based on a YOLOv8 model, fuses a Bottleneck Transformer into a bottleneck layer of C2f to obtain a C2f_Bottleneck_BoT3 module, and applies the C2f_Bottleneck_BoT3 module to a neck network of the YOLOv8; The improved YOLOv8 model uses a cavity convolution module as a spatial pyramid pooling structure connecting a backbone network and the neck network of the YOLOv8 model; inputting the weld image to be tested into the trained weld position segmentation model to output initial weld position feature information; Mask processing the initial weld position feature information to obtain an initial weld position mask binary image; The mask processing of the initial weld position feature information comprises the following steps: Determining a color value interval of the weld mask through an H component histogram; Setting the value range of the H component to 0-25, then fine-tuning the S component and the V component, and generating a mask binary image of the initial weld position feature of the medium-thick plate according to the color limit of the HSV model; The noise in the generated mask binary image is removed by erosion and dilation.

2. The BoT3-YOLOv8-based plate initial welding position segmentation method according to claim 1, characterized in that: The C2f_Bottleneck_BoT3 module adopts a multi-head self-attention mechanism.

3. The BoT3-YOLOv8-based plate initial welding position segmentation method according to claim 1, characterized in that: For a given size of image input, the cavity convolution module uses different sampling rates of cavity convolution to sample the input image, then splices the output results of each cavity convolution at the head and tail, expands the channel number of the image, and finally reduces the channel number through 1x1 convolution to output the image.

4. A medium plate initial welding position segmentation system based on BoT3-YOLOv8, characterized by: The system comprises a collection unit, a preprocessing unit, a model training unit, a weld position segmentation unit, and a weld position mask unit; The collection unit is configured to collect weld image data of a medium-thick plate; The preprocessing unit is configured to preprocess the weld image data to obtain processed image data; The model training unit is configured to train a weld position segmentation model using the processed image data to obtain a trained weld position segmentation model; the weld position segmentation model comprises an improved YOLOv8 model; The improved YOLOv8 model is based on a YOLOv8 model, fuses a Bottleneck Transformer into a bottleneck layer of C2f to obtain a C2f_Bottleneck_BoT3 module, and applies the C2f_Bottleneck_BoT3 module to a neck network of the YOLOv8; The improved YOLOv8 model uses a cavity convolution module as a spatial pyramid pooling structure connecting a backbone network and the neck network of the YOLOv8 model; The welding position segmentation unit is configured to input a welding seam image to be detected into the trained welding position segmentation model, and output initial welding position feature information; The welding position mask unit is configured to perform mask processing on the initial welding position feature information to obtain an initial welding position mask binary image. The initial welding position feature information is subjected to mask processing by using an HSV color space region segmentation method, and the mask processing specifically includes: The color value interval of the welding seam mask is determined through an H component histogram; The value range of the H component is set to 0-25, and then the S component and the V component are fine-tuned, and a mask binary image of the initial welding position feature of the medium plate is generated according to the color limit of the HSV model; The generated mask binary image is subjected to denoising by using an erosion and dilation processing method.

5. The BoT3-YOLOv8-based plate initial welding position segmentation system according to claim 4, characterized in that: The C2f_Bottleneck_BoT3 module adopts a multi-head self-attention mechanism.

Images