Workpiece size measuring method based on improved YOLO segmentation algorithm and carrying rivet welding device

Abstract

The invention provides a workpiece size measuring method based on an improved YOLO segmentation algorithm and a carrying rivet welding device. The workpiece type is recognized by using the improved YOLO segmentation algorithm and a binocular depth camera, a workpiece image is segmented, and the size of a workpiece and the spatial position of workpiece grabbing are determined; and carrying and rivet welding tasks are executed according to the space coordinates of the grabbing positions of the workpieces with the needed sizes, and after carrying and rivet welding work is finished, automatic welding work is completed. According to the device, the workpiece type can be automatically recognized, the workpiece size can be measured, and the workpiece grabbing space position can be determined, so that the whole process of carrying, rivet welding and welding is automatic, and the problems that the automation degree is low, the production efficiency is low, the manpower consumption is large and potential safety hazards exist in the traditional manual carrying and rivet welding process are solved to a certain extent; and the intelligence and integration degree of the robot welding field is improved, and the welding quality is further guaranteed.

Description

Technical Field

[0001] The present application relates to the field of robotics technology, and in particular to a workpiece size measurement method and a handling riveting and welding device based on an improved YOLO segmentation algorithm. Background Art

[0002] Welding robots are industrial automation equipment used to automatically perform welding operations. They are widely used in many fields such as automobile manufacturing, shipbuilding, aerospace, metal processing, etc., and have brought great changes and progress to modern industrial production. With the continuous development of industrial automation, they are also constantly developing and improving, and combined with other intelligent technologies, they are driving the manufacturing industry towards a more intelligent and unmanned process.

[0003] As we all know, the automation of the entire process of welding robots can greatly improve industrial production efficiency and safety. At present, welding robots have basically realized manual riveting and automatic welding of assembled workpieces, while the workpiece handling, riveting and assembly processes before welding are completely completed manually. The angle and relative position of the riveting between the workpieces are roughly estimated by the mold, and the position and angle of the workpiece to be welded on the robot positioner platform also rely on the workers' experience. These extensive production processes greatly affect the welding accuracy in the subsequent automated welding process, making it difficult to achieve welds with higher levels of requirements, thereby affecting the welding quality of the final product. This traditional manual handling and riveting process has a low degree of automation, low production efficiency, high labor consumption and certain safety hazards, which is seriously inconsistent with the requirements of my country's industrial production modernization construction. At the same time, the current algorithms for classification detection, image segmentation and size measurement of specific target types are often limited by the difficulties of insufficient network feature extraction capabilities, low classification and segmentation robustness, and visual measurement accuracy that is easily affected by ambient light, and are difficult to deploy in industrial sites, further restricting the development of industrial intelligence. Summary of the invention

[0004] In order to solve the above technical problems, the present invention discloses a workpiece size measurement method and a handling riveting welding device based on an improved YOLO segmentation algorithm, which improves the automation and intelligence level in the field of robot welding.

[0005] Specifically, the present application provides a workpiece size measurement method based on an improved YOLO segmentation algorithm, comprising the following steps:

[0006] S1: After initialization, start the binocular depth camera to collect images of the workpiece area to be riveted;

[0007] S2: Use the workpiece detection system to identify the workpiece type, segment the workpiece image and measure the workpiece size to determine whether the transported object is a workpiece of the required type and correct size;

[0008] S3: Obtain the three-dimensional coordinates of the center point of the target workpiece as the grasping position of the handling robot, and transmit the coordinate information to the handling robot controller;

[0009] S4: The transport robot moves to the designated grabbing position to grab the workpiece, and transports the workpiece to the pre-taught position;

[0010] S5: After the workpiece is moved to the teaching position, the welding robot performs riveting welding of the workpiece;

[0011] S6: After the riveting and welding of the current workpiece is completed, the processes from S1 to S5 are carried out in sequence for the subsequent workpieces to be riveted until the riveting and welding assembly of the entire product is completed.

[0012] The initialization in S1 specifically includes: calibrating the binocular depth camera, calibrating the hand-eye of the binocular depth camera and the robot, and loading a pre-established mathematical model of the workpiece to be welded;

[0013] Furthermore, the initialization in S1 also includes: the transport and assembly sequence of the workpiece, the target position of the workpiece transport, the position of the workpiece riveting point and the preset parameters of the mathematical model of the welding workpiece, and the preset parameters at least include the workpiece type and the workpiece size, but are not limited thereto.

[0014] Wherein, the workpiece detection system in S2 includes an image segmentation module and a workpiece size measurement module;

[0015] Furthermore, the image segmentation module in S2 uses an improved YOLO segmentation algorithm to perform workpiece type recognition and workpiece image segmentation on the collected image data;

[0016] The workpiece size measurement module in S2 uses the workpiece spatial position information obtained by the binocular depth camera to measure the size of the target workpiece and determine the three-dimensional coordinates of the required workpiece grasping position.

[0017] In the above technical solution, the improved YOLO segmentation algorithm in S2 includes the SDEYolo segmentation algorithm, which is used for workpiece type recognition and image segmentation;

[0018] The SDEYolo segmentation algorithm in S2 includes: an input end, a Backbone network, a Neck network, and a Head network; and also includes:

[0019] Add an EMA attention mechanism based on efficient multi-scale without dimensionality reduction between the feature fusion convolution module C2f and the Conv standard convolution layer of the Backbone network;

[0020] Replace the feature fusion module in the Neck network with the SDI multi-level feature fusion module;

[0021] Replace the Conv standard convolution layer of the Neck network with a switchable hole convolution module SAConv;

[0022] In the above technical solution, the SDEYolo segmentation algorithm in S2 further includes: using a diversified branch block DBB to improve the segmentation detection head segment in the Head network to form a new segmentation detection head DBB_segment;

[0023] The diversified branch block DBB in the SDEYolo segmentation algorithm uses branches of different scales and complexities and convolution kernels and average pooling layers of different sizes. At the same time, a complex branch structure is used in the training stage, and in the inference stage, these branches are equivalently converted into a single convolution layer to maintain efficient inference;

[0024] In the above technical solution, the feature fusion convolution module C2f in the SDEYolo segmentation algorithm also includes: 2 Conv standard convolution layers, 1 Split group feature extraction module and n Bottleneck modules.

[0025] Based on the same concept, the present application also provides a transport riveting and welding device, which comprises at least: a transport six-axis robot, a welding six-axis robot, a robot controller, a welder, a robot welding positioner, a binocular depth camera and a host computer;

[0026] The robot controller, binocular depth camera, and host computer cooperate to control the six-axis handling robot to achieve workpiece size measurement and handling; the robot controller controls the six-axis welding robot, robot welding positioner, and welding machine to perform automatic riveting and welding;

[0027] The host computer integrates the SDEYolo segmentation algorithm in the workpiece size measurement method based on the improved YOLO segmentation algorithm, including an image segmentation module and a workpiece size measurement module, which are used for image detection, segmentation, size measurement and determination of the spatial coordinates of the workpiece grasping position;

[0028] Wherein, the host computer, robot controller and welding machine communicate via TCP / IP to transmit signals;

[0029] The binocular depth camera is used to capture the image of the workpiece to be riveted and welded, and is connected to the host computer for communication to complete the image and spatial position information acquisition of the workpiece to be riveted and welded.

[0030] In summary, the present invention provides a workpiece size measurement method and a handling and riveting welding device based on an improved YOLO segmentation algorithm. The improved YOLO segmentation algorithm is used to identify the workpiece type, segment the workpiece image, determine the size of the workpiece and the spatial position of the workpiece grasping; according to the grasping position of the target workpiece, the handling and riveting welding procedures are executed, and after the handling and riveting welding are completed, the laser scanning automatic welding procedure is executed.

[0031] Compared with the prior art, the present invention realizes automatic identification of workpiece type, measurement of workpiece size and determination of the spatial position of workpiece grabbing by using an improved YOLO segmentation algorithm, thereby realizing automation of the entire process of handling, riveting and welding. To a certain extent, it overcomes the difficulties of low automation, low production efficiency, high manpower consumption and safety hazards in traditional manual handling and riveting processes, improves the intelligence and integration level in the field of robot welding, and further ensures the welding quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flowchart of the workpiece size measurement method based on the improved YOLO segmentation algorithm described in this application.

[0033] Figure 2 It is a structural block diagram of the SDEYolo segmentation algorithm described in the present invention.

[0034] Figure 3 This is a structural block diagram of the SAConv switchable dilated convolution module described in the present invention, which includes two global context modules and a SAC convolution module.

[0035] Figure 4 This is a block diagram of the DBB diversified branching block structure of the present invention, which includes a training phase DBB module and an inference phase DBB module.

[0036] Figure 5 It is a module schematic diagram of an embodiment of a handling riveting and welding device based on an improved YOLO segmentation algorithm according to the present invention, wherein a welding six-axis robot 1, a handling six-axis robot 2, a binocular depth camera 3, a robot welding positioner 4, a robot controller 5, a host computer 6 and a welding machine 7. DETAILED DESCRIPTION

[0037] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] Embodiment 1:

[0039] See also Figure 1 The present application provides a workpiece size measurement method based on an improved YOLO segmentation algorithm, which realizes automatic identification of workpiece type, segmentation of workpiece image, and determination of workpiece size, thereby improving the degree of automation and intelligence in the field of robot welding.

[0040] Specifically, the workpiece size measuring method comprises the following steps:

[0041] S1: After initialization, start the binocular depth camera to collect images of the workpiece area to be riveted;

[0042] S2: Use the workpiece detection system to identify the workpiece type, segment the workpiece image and measure the workpiece size to determine whether the transported object is a workpiece of the required type and correct size;

[0043] S3: Obtain the three-dimensional coordinates of the center point of the target workpiece as the grasping position of the handling robot, and transmit the coordinate information to the handling robot controller;

[0044] S4: The transport robot moves to the designated grabbing position to grab the workpiece, and transports the workpiece to the pre-taught position;

[0045] S5: After the workpiece is moved to the teaching position, the welding robot performs riveting welding of the workpiece;

[0046] S6: After the riveting and welding of the current workpiece is completed, the processes from S1 to S5 are carried out in sequence for the subsequent workpieces to be riveted until the riveting and welding assembly of the entire product is completed.

[0047] Wherein, the workpiece size measurement system in S2 includes an image segmentation module and a workpiece size measurement module;

[0048] The image segmentation module in S2 uses an improved YOLO segmentation algorithm to perform workpiece type recognition and workpiece image segmentation on the collected image data;

[0049] The workpiece size measurement module in S2 uses the workpiece spatial position information obtained by the binocular depth camera to measure the size of the target workpiece and determine the three-dimensional coordinates of the required workpiece grasping position.

[0050] The SDEYolo segmentation algorithm in S2 includes:

[0051] On the input side, mosaic data augmentation and cutting and pasting of partial images are used to enrich the dataset, so that when the captured images are of different sizes, adaptive image scaling can be performed, which reduces the amount of calculation and improves the detection speed.

[0052] The Backbone network is responsible for extracting features from the input image. It adopts a variant based on the advanced CSPDarknet network architecture and uses the C2f (CSPDarknet53 to 2-StageFPN) module to replace the previous version C3 module. While maintaining better performance, it has a lower computational cost, making the backbone network more efficient and powerful.

[0053] The Neck network aggregates features from multiple levels of the backbone, including two upsampling and four concat operations, to enhance the fusion of features of different scales, enabling the model to better detect objects of various sizes.

[0054] The Head network provides efficient and accurate target detection and segmentation capabilities through dense prediction, anchor-free mechanism, advanced regression and classification methods, multi-scale prediction, fusion and branching structure, and non-maximum suppression.

[0055] like Figure 2-4 As shown, the SDEYolo segmentation algorithm in S2 also includes the following improvements:

[0056] (1) An EMA attention mechanism based on efficient multi-scale without dimensionality reduction is added between the feature fusion convolution module C2f and the Conv standard convolution layer of the Backbone network.

[0057] In the EMA attention module, first, the input is grouped in the channel dimension, and then processed differently by different branches, one branch performs one-dimensional global pooling, and the other branch performs feature extraction through 3×3 convolution. The output features of the two branches are then modulated by the sigmoid function and normalization operation, and then merged through the cross-dimensional interaction module to capture the pairwise relationship at the pixel level. Finally, after being adjusted by the sigmoid function, the output feature map is used to enhance or weaken the original input features to obtain the final output.

[0058] By introducing the EMA attention mechanism based on efficient multi-scale without dimensionality reduction, the network can capture pixel-level relationships by using cross-dimensional interactions. At the same time, this mechanism retains the key information of each channel at a very low computational cost, thereby enhancing the model's ability to process features and the effect of feature representation, thereby improving the robustness of image segmentation.

[0059] (2) The feature fusion module in the Neck network is replaced with the SDI multi-level feature fusion module to enhance the integration of semantic information into low-level features, and finer details are used to improve high-level features, thereby improving the performance of image segmentation.

[0060] In the SDI multi-level feature fusion module, SDI first performs channel compression and spatial expansion on the feature map inputs from different levels and channels, so that the feature maps of different levels or channels have the same dimension. Further, SDI applies spatial and channel attention mechanisms to process the features of each level, so that the features can integrate local spatial information and global channel information to better capture the key information in the image. Finally, for the feature maps of each level, SDI uses the Hadamard product operation to fuse high-level features containing more semantic information and low-level features that can capture finer details, thereby enhancing the semantics and details of the features of each level. After these operations, SDI outputs the fused feature maps for subsequent operations.

[0061] By replacing the feature fusion module with the SDI multi-level feature fusion module, the model can greatly improve the image segmentation accuracy without significantly increasing the computational complexity of the model.

[0062] (3) Replace the Conv standard convolution layer of the Neck network with a switchable hole convolution module SAConv.

[0063] The switchable dilated convolution module SAConv can be subdivided into two global context modules and one SAC convolution module. In this structure, each global context module consists of an average pooling layer and a 1×1 convolution layer, which are placed at the front and back ends of the SAC convolution module. Their role is to add global information to the feature map and enhance the network's comprehensive understanding of the image content, so that the SAC convolution module can operate efficiently in a wider context and significantly improve the quality and accuracy of feature extraction. Inside the SAC convolution module, first, the module applies convolution operations with different dilation rates to the input features to capture multi-scale feature information without increasing parameters or computational burden. The so-called dilation refers to the introduction of additional space in the convolution kernel to expand the receptive field. Then, the SAC convolution module uses a space-dependent switching function to integrate the results of convolutions with different dilation rates, so that the network can more flexibly process features of different sizes and scales. This process is achieved by sharing weights and using the same trainable weights (except for a difference term) in convolutions with different dilation rates. By integrating innovative designs and mechanisms such as global context modules, switch functions, multi-void convolutions, and shared weights, the SAConv module can significantly enhance the adaptability and accuracy of the network when processing features of various scales and complexities, thereby enabling SAConv to demonstrate outstanding performance improvements in application areas such as object detection and image segmentation.

[0064] (4) The diversified branch block DBB is used to improve the segmentation detection head segment in the Head network to form a new segmentation detection head DBB_segment.

[0065] The specific implementation principle of the segmentation detection head DBB_segment in the SDEYolo segmentation algorithm is:

[0066] Module replacement: The DBB module contains multiple branches of different types, each of which contains convolution kernels of different sizes (such as 3×3, 5×5) and different pooling layers (such as average pooling, maximum pooling). These branches work together to capture different features of the input data. The specific operation is to find the appropriate convolution layer position in the original segmentation head of YOLO, replace these convolution layers with diversified branch blocks DBB to complete the new segmentation detection head DBB_segment, for example, add a diversified branch block DBB after the upsampling layer or introduce a diversified branch block DBB in the middle layer of the segmentation head, but not limited to this.

[0067] Model training: Use the modified segmentation detection head DBB_segment containing DBB for training. During the training phase, the outputs of each branch are fused together through concatenation or weighted summation to generate a diverse feature representation, ensuring that the multi-branch structure can work properly during training and optimizing model parameters.

[0068] Merge operation: In the inference stage, the multi-branch structure is equivalently merged into a single convolution by merging the convolution kernel and the bias term, realizing the merging of the multi-branch structure into a single equivalent convolution layer, reducing the computational overhead and enabling the model to maintain high inference efficiency during inference.

[0069] In the above technical solution, the feature fusion convolution module C2f in the SDEYolo segmentation algorithm also includes: 2 Conv standard convolution layers, 1 Split grouping feature extraction module and n Bottleneck modules, wherein:

[0070] Conv standard convolution layer: Each Conv standard convolution layer includes a fixed-size convolution kernel, a batch normalization layer, and an activation function, which are used to extract and process features to ensure the stability and generalization ability of the network during training.

[0071] Bottleneck module: Each Bottleneck module consists of a 1×1 Conv standard convolution layer, a 3×3 Conv standard convolution layer, and a residual connection operation. Among them, the 1×1 Conv standard convolution layer is used to adjust the number of channels to reduce or increase the number of channels of the feature map, and the 3×3 Conv standard convolution layer is used to extract local features, capture spatial information, and add the input directly to the output through the residual connection operation to achieve direct feature transmission and smooth flow of gradients, alleviate the gradient disappearance problem, and enhance the training effect of the network.

[0072] Split group feature extraction module: The Split module splits the input feature map into several sub-feature maps in the channel dimension. Through segmentation, parallel processing of feature maps is achieved, improving computational efficiency and flexibility of feature extraction.

[0073] By applying 2 Conv standard convolutional layers, 1 Split group feature extraction module and several Bottleneck modules, the C2f module achieves efficient feature extraction and enhances information flow, thereby significantly improving the performance of the model in object detection and segmentation tasks.

[0074] Embodiment 2:

[0075] like Figure 5 As shown, the present application also provides a transport riveting and welding device, which at least includes: a transport six-axis robot, a welding six-axis robot, a robot controller, a welder, a robot welding positioner, a binocular depth camera and a host computer;

[0076] The robot controller, binocular depth camera, and host computer cooperate to control the six-axis handling robot to achieve workpiece size measurement and handling; the robot controller cooperates to control the six-axis welding robot, the robot welding positioner, and the welding machine to perform automatic riveting and welding; the robot controller controls the movement of the six-axis robot, including movement in the joint coordinate system and movement in the world coordinate system;

[0077] The host computer incorporates the SDEYolo segmentation algorithm in the workpiece size measurement method based on the improved YOLO segmentation algorithm, including an image segmentation module and a workpiece size measurement module, which are used for image detection, segmentation, size measurement and determination of the spatial coordinates of the grasping position of the workpiece; the handling robot grasps and carries the workpiece in sequence according to the transmitted spatial coordinates of the grasping position;

[0078] Wherein, the host computer, robot controller and welding machine communicate via TCP / IP to transmit signals;

[0079] The binocular depth camera is used to capture the image of the workpiece to be riveted and connected to the host computer to complete the acquisition of the regional image and spatial coordinate information of the workpiece to be riveted; the binocular depth camera and the host computer generally communicate via USB, but not limited to this.

[0080] The working process of a handling riveting and welding device of the present invention is as follows: (1) the binocular depth camera and the laser are calibrated respectively, the binocular depth camera and the handling robot are calibrated by hand and eye, the laser and the welding robot are calibrated by hand and eye, and a pre-established mathematical model of the workpiece to be welded is loaded; (2) the operator determines the mathematical model of the welding workpiece and its preset parameters, the handling and assembly sequence of the workpiece, the target position of the workpiece handling, and the position of the workpiece riveting point; (3) taking the tower foot as an example, after the above operations are completed, the binocular depth camera is started to start collecting images of the workpiece area to be riveted and welded, and the upper computer takes the images taken by the binocular depth camera and obtains the images obtained by the workpiece size measurement system. The spatial coordinates of the center point of the workpiece of the required size are sent to the robot controller. The robot controller generates motion instructions based on the spatial coordinates of the center point of the workpiece of the required size, the target position of the workpiece transportation, and the position of the workpiece riveting point, etc. to drive the robot arm to grab the base plate workpiece and transport it to the positioner, and then perform the same operation to grab the main plate workpiece and transport it to the positioner; (4) Further, start the welding robot to perform automatic riveting and welding. After the first riveting and welding is completed, the handling robot performs the same grabbing operation as mentioned above, and performs handling and riveting and welding according to the preset assembly sequence; (5) After the handling and riveting are all completed, the welding robot executes the automatic welding program to complete the laser scanning automatic welding.

[0081] Although example embodiments have been described herein with reference to the accompanying drawings, it should be understood that the above example embodiments are merely exemplary and are not intended to limit the scope of the present application to this. Those of ordinary skill in the art may make various changes and modifications therein without departing from the scope and spirit of the present application. All these changes and modifications are intended to be included within the scope of the present application as required by the appended claims.

[0082] It should be noted that, in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is a relationship between these entities or operations.

[0083] In any such actual relationship or order. Moreover, the terms "comprises," "comprising," or any other variation thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a list of elements includes not only those elements, but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus comprising the element.

[0084] Although the description of the present application is carried out in conjunction with the above specific embodiments, it is obvious that those skilled in the art can make many substitutions, modifications and changes based on the above content. Therefore, all such substitutions, improvements and changes are included in the spirit and scope of the appended claims.

Claims

1. A method for measuring the size of a workpiece based on an improved YOLO segmentation algorithm for a riveting and welding device, characterized in that: The method comprises: S1: After initialization, start the binocular depth camera to collect images of the workpiece area to be riveted; S2: Use the workpiece detection system to identify the workpiece type, segment the workpiece image and measure the workpiece size to determine whether the transported object is a workpiece of the required type and correct size; S3: Obtain the three-dimensional coordinates of the center point of the target workpiece as the grasping position of the handling robot, and transmit the coordinate information to the handling robot controller; S4: The transport robot moves to the designated grabbing position to grab the workpiece, and transports the workpiece to the pre-taught position; S5: After the workpiece is moved to the teaching position, the welding robot performs riveting welding of the workpiece; S6: After the riveting and welding of the current workpiece is completed, the processes from S1 to S5 are carried out in sequence for the subsequent workpieces to be riveted until the riveting and welding assembly of the entire product is completed.

2. According to the workpiece size measurement method based on the improved YOLO segmentation algorithm according to claim 1, it is characterized in that: The initialization in S1 specifically includes: calibrating the binocular depth camera, performing hand-eye calibration on the binocular depth camera and the robot, loading a pre-established mathematical model of the workpiece to be welded, determining the workpiece handling and assembly sequence, the target position of the workpiece handling, the position of the workpiece riveting point, and the preset parameters of the mathematical model of the welding workpiece, wherein the preset parameters include at least the workpiece type and the workpiece size.

3. The workpiece size measurement method based on the improved YOLO segmentation algorithm according to claim 2 is characterized in that: The workpiece detection system in S2 includes an image segmentation module and a workpiece size measurement module, specifically: The image segmentation module uses an improved YOLO segmentation algorithm to perform workpiece type recognition and workpiece image segmentation on the collected image data; The workpiece size measurement module uses the workpiece spatial position information obtained by the binocular depth camera to measure the size of the target workpiece and determine the three-dimensional coordinates of the required workpiece grasping position.

4. The workpiece size measurement method based on the improved YOLO segmentation algorithm according to claim 3 is characterized in that: The improved YOLO segmentation algorithm in S2 includes the SDEYolo segmentation algorithm, which is used for artifact type detection and image segmentation; The SDEYolo segmentation algorithm in S2 includes: an input end, a Backbone network, a Neck network, and a Head network; The specific improvements are: adding an EMA attention mechanism based on efficient multi-scale without dimensionality reduction between the feature fusion convolution module C2f and the Conv standard convolution layer of the Backbone network; Replace the feature fusion module in the Neck network with the SDI multi-level feature fusion module; The convolution module of the Neck network is replaced with a switchable hole convolution module SAConv.

5. The workpiece size measurement method based on the improved YOLO segmentation algorithm according to claim 4 is characterized in that: The SDEYolo segmentation algorithm in S2 is specifically: A diversified branch block DBB is used to improve the segmentation detection head segment in the Head network to form a new segmentation detection head DBB_segment; the diversified branch block DBB uses branches of different scales and complexities, such as convolution kernels of different sizes and average pooling layers. At the same time, a complex branch structure is adopted in the training stage, and in the reasoning stage, these branches are equivalently converted into a single convolution layer to maintain efficient reasoning.

6. A workpiece size measurement method based on an improved YOLO segmentation algorithm according to any one of claims 4-5, characterized in that: The C2f feature fusion convolution module specifically includes: 2 Conv standard convolution layers, 1 Split grouping feature extraction module and n Bottleneck modules.

7. A handling riveting and welding device for a workpiece size measurement method based on an improved YOLO segmentation algorithm according to any one of claims 1 to 6, characterized in that: The transport riveting and welding device at least comprises: a transport six-axis robot, a welding six-axis robot, a robot controller, a welder, a robot welding positioner, a binocular depth camera and a host computer.

8. The transport riveting and welding device according to claim 7, characterized in that: Specifically, the robot controller, binocular depth camera, and host computer cooperate to control the six-axis handling robot to achieve workpiece size measurement and handling; The robot controller controls the six-axis welding robot, the robot welding positioner and the welding machine to perform automatic riveting and welding; The host computer integrates the SDEYolo segmentation algorithm in the workpiece size measurement method based on the improved YOLO segmentation algorithm, including an image segmentation module and a workpiece size measurement module, which are used for image detection, segmentation, size measurement and determining the spatial coordinates of the workpiece grasping position.

9. The transport riveting and welding device according to claim 8, characterized in that: The host computer, robot controller and welding machine communicate via TCP / IP to transmit signals; The binocular depth camera is used to capture the image of the workpiece to be riveted and welded, and is connected to the host computer for communication to complete the image and spatial position information acquisition of the workpiece to be riveted and welded.

Images