3D vision measurement welding method based on deep learning

The 3D visual measurement welding method improves weld defect recognition accuracy to 99% by employing deep learning techniques like Faster-RCNN, RPN, and FPN, enhancing small target detection and reducing computational costs.

CN119057291APending Publication Date: 2024-12-03SIPUQIAN (SUZHOU) INTELLIGENT EQUIPMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411247324.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-06
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing 3D vision measurement technologies for weld detection face challenges in accurately recognizing small targets due to complex shape and scale changes in weld images, leading to low accuracy and precision in weld defect recognition.

Method used

A 3D visual measurement welding method using deep learning, incorporating Faster-RCNN with RPN and RoIPooling, ResNet, and a multi-scale feature fusion mechanism (FPN), along with transfer learning and model fine-tuning, to enhance weld feature extraction and detection accuracy.

Benefits of technology

The method achieves 99% accuracy in weld defect recognition, with an 8% increase in accuracy, 5% increase in precision, and 7% increase in recall rate compared to the original Faster R-CNN model, addressing the issue of small target detection loss.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119057291A_ABST
    Figure CN119057291A_ABST
Patent Text Reader

Abstract

The invention discloses a 3D vision measurement welding method based on deep learning, and relates to the field of robot welding, and the method comprises the following steps: step 1, image measurement preparation: before welding is started, a 3D vision system scans a welding seam through a high-precision 3D camera to obtain three-dimensional point cloud data of the welding seam, and the three-dimensional point cloud data of the welding seam are obtained through an intelligent algorithm; the system can extract accurate position information of the welding seam; in the welding process, a 3D visual system can monitor the position and state of a welding seam in real time, a welding track is dynamically adjusted, and it is ensured that a welding gun or a welding robot can accurately track the welding seam for welding; and 3, detection and evaluation are conducted, specifically, after welding is completed, a 3D visual system can conduct quality detection on a welding seam, and various indexes are evaluated by comparing three-dimensional point cloud data before and after welding. According to the method, the problem of small target detection loss in the image conversion feature extraction process is solved, and the accuracy is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention mainly relates to the technical field of robot welding, and specifically to a 3D vision measurement welding method based on deep learning. Background Art

[0002] Faster-RCNN weld feature extraction technology is a deep learning technology widely used in weld detection. It combines the advantages of convolutional neural network (CNN) and region proposal network (RPN), and can efficiently extract features in weld images for defect detection.

[0003] 3D vision measurement technology is an advanced high-end technology that combines computer vision and precision measurement. It imitates the way human eyes observe the world and uses computer analysis and processing to perceive the surrounding environment to identify, locate, quantify and reconstruct the observed objects.

[0004] Due to the complex shape, scale changes and different levels of features of weld images, it is very easy to cause the loss of detection of small targets, especially in the process of image conversion feature extraction, so that the final weld defect recognition accuracy cannot reach a higher level, and the accuracy and precision rate are relatively low. Summary of the invention

[0005] Based on this, the purpose of the present invention is to provide a 3D vision measurement welding method based on deep learning to solve the technical problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A 3D visual measurement welding method based on deep learning, the visual measurement welding method comprising the following steps:

[0008] Step 1: Image measurement preparation: Before welding begins, the 3D vision system scans the weld with a high-precision 3D camera to obtain the 3D point cloud data of the weld. Using intelligent algorithms, the system can extract the accurate position information of the weld, including the geometric shape, size and position coordinates of the weld;

[0009] Step 2: Image preprocessing: During the welding process, the 3D vision system can monitor the position and status of the weld in real time and dynamically adjust the welding trajectory to ensure that the welding gun or welding robot can accurately track the weld for welding;

[0010] Step 3: Inspection and evaluation: After welding is completed, the 3D vision system can perform quality inspection on the weld. By comparing the 3D point cloud data before and after welding, the key indicators of the weld geometry, dimensional accuracy and surface quality can be evaluated.

[0011] Preferably, in steps one to three, based on the weld extraction technology of deep learning (Faster-RCNN), the region proposal network (RPN) and candidate region pooling (RoIPooling) layer are introduced into the Faster-RCNN classic target detection deep neural network in view of the complex shape, scale changes and features of different levels of the weld image.

[0012] Preferably, the candidate region pooling (RoIPooling) layer adopts a residual neural network (ResNet), which uses residual learning to connect deep feature maps with previous shallow feature maps, effectively utilizing high-level and low-level features to adapt to the detection of small targets;

[0013] The scale changes and features at different levels in weld images are taken into account. A multi-scale feature fusion mechanism, FPN, is used. By introducing side connections and upsampling operations in the feature extraction network, feature maps from different levels are fused to generate a feature pyramid with multi-scale information.

[0014] Preferably, in the region proposal network (RPN), the improved model adjusts the size and proportion of the anchor box, adopts different sliding window strategies or introduces a more effective candidate region generation mechanism according to the characteristics of the weld image, so as to improve the accuracy and recall rate of the candidate region;

[0015] Classification loss function and bounding box regression loss for weld targets, weighted, balanced or multi-task learning strategies for weld features to improve the performance and robustness of the model;

[0016] The classification loss function is as follows:

[0017]

[0018] in,

[0019]

[0020] Preferably, the deep learning (Faster-RCNN) weld extraction technology also includes transfer learning and model fine-tuning;

[0021] The transfer learning and model fine-tuning perform transfer learning based on a pre-trained model, and perform model fine-tuning for a weld feature extraction task; by performing fine-tuning on a weld dataset, the model can better adapt to the characteristics of weld images, thereby improving detection performance and generalization capabilities.

[0022] Preferably, during the transfer learning, the following transfer loss function based on optimal transmission is used to optimize the model;

[0023] min π∈∏(μ,ν) ∫x×y c(x,y)dπ(x,y).

[0024] Preferably, the 3D vision in steps 1 to 3 is a 3D vision measurement technology based on digital micromirror technology (DLP) DLP projection, and according to the parameter information of the structured light system equipment (such as resolution and grayscale range), a suitable coefficient is set to design a sinusoidal variation function to generate a grayscale stripe coding sequence; then the obtained grayscale stripe coding sequence is assigned to each row of the coding pattern, that is, the grayscale coding information of each row of the coding pattern is the same; finally, the stripes are colored with a color to obtain a monochrome sinusoidal stripe coding pattern;

[0025] All stripes in the coding pattern are coded in red, and there are black stripes between the red stripes; in addition to the different stripe color information, there is also a difference in the coding of grayscale values ​​in the horizontal direction between the proposed coding pattern and the color stripe coding pattern;

[0026] The color stripe coding pattern uses binary coding, and the proposed coding pattern uses multi-value coding and satisfies sinusoidal variation.

[0027] Preferably, based on the monochrome sinusoidal stripes, all stripes are encoded using the same color, and random points are embedded around the stripes, and the random points are used to identify the stripes based on the distribution of random points around the stripes; finally, the generated pseudo-random dot pattern and the proposed monochrome sinusoidal stripe coding pattern are combined together to provide coding information for the identification of monochrome stripes.

[0028] Preferably, in order to make the random dots highly distinguishable in the encoded structured light pattern and the captured structured light image, regular restrictions are imposed on the random dot size, random dot density and random dot distribution.

[0029] Preferably, the coding pattern only requires two colors, and the two colors can be reconstructed and transformed according to the colored object.

[0030] In summary, the present invention mainly has the following beneficial effects:

[0031] This project developed a weld feature extraction model based on the improved Faster-RCNN. By introducing the region proposal network and the candidate region pooling layer, key features were extracted from the weld image and the weld position was detected. The accuracy of weld defect recognition reached 99%. In particular, the industry problem of small target detection loss was solved. The application of this technology has achieved an accuracy of 99% in the recognition of different types of weld defects. Compared with the original Faster R-CNN model, the accuracy rate increased by 8%, the precision rate increased by 5%, and the recall rate increased by 7%, which improved the accuracy of weld detection in the industry.

[0032] The 3D vision measurement technology based on DLP projection is developed. Through DLP projection coding patterns, monochrome sinusoidal stripe patterns and pseudo-random dot patterns are developed and projected. The designed coding pattern only requires two colors, and the coding color is changed according to the color of the object being measured. The stripes are identified by pseudo-random points, and the stripe decoding is independent of each other, reducing the decoding errors caused by abnormal points, and thus achieving dense three-dimensional reconstruction, solving the problem of reflective surface of welded objects and difficulty in three-dimensional reconstruction. The success rate of one-time projection reconstruction is 99.8%. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is the overall process framework diagram of the present invention;

[0034] Figure 2 It is a pyramid network structure diagram of the FPN of the present invention;

[0035] Figure 3 This is a comparison diagram of the actual image and the predicted image of the weld identification detection of the present invention;

[0036] Figure 4 This is a comparison diagram of the actual image and the predicted image of the weld identification detection of the present invention;

[0037] Figure 5 This is a comparison diagram of the actual image and the predicted image of the weld identification detection of the present invention;

[0038] Figure 6 This is a comparison diagram of the actual image and the predicted image of the weld identification detection of the present invention;

[0039] Figure 7 This is a comparison diagram of the actual image and the predicted image of the weld identification detection of the present invention;

[0040] Figure 8 It is a test diagram for comparing the accuracy of the real image and the predicted image of the weld identification detection of the present invention;

[0041] Fig. 9 It is a diagram of the monochrome stripe pattern and the gray value change along the horizontal direction in the sinusoidal stripe pattern of the present invention. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0043] Example

[0044] like Figure 1 As shown, the visual measurement welding method includes the following steps:

[0045] Step 1: Image measurement preparation: Before welding begins, the 3D vision system scans the weld with a high-precision 3D camera to obtain the 3D point cloud data of the weld. Using intelligent algorithms, the system can extract the accurate position information of the weld, including the geometric shape, size and position coordinates of the weld;

[0046] Step 2: Image preprocessing: During the welding process, the 3D vision system can monitor the position and status of the weld in real time and dynamically adjust the welding trajectory to ensure that the welding gun or welding robot can accurately track the weld for welding;

[0047] Step 3: Inspection and evaluation: After welding is completed, the 3D vision system can perform quality inspection on the weld. By comparing the 3D point cloud data before and after welding, the key indicators of the weld geometry, dimensional accuracy and surface quality can be evaluated.

[0048] It should be noted that, in this embodiment, in the pre-welding preparation stage, weld identification and positioning: before welding begins, the 3D vision system scans the weld with a high-precision 3D camera to obtain the three-dimensional point cloud data of the weld; using intelligent algorithms, the system can extract the accurate position information of the weld, including the geometric shape, size and position coordinates of the weld; workpiece alignment and positioning: in addition to weld identification, the 3D vision system can also perform overall alignment and positioning of the workpiece to ensure the position accuracy of the workpiece before welding and reduce welding quality problems caused by workpiece position deviation;

[0049] Real-time weld tracking during welding: During welding, the 3D vision system can monitor the position and status of the weld in real time and dynamically adjust the welding trajectory to ensure that the welding gun or welding robot can accurately track the weld for welding; Welding posture adjustment: Based on the real-time weld information, the system can calculate the optimal welding posture of the welding robot or welding equipment, including parameters such as welding angle, welding speed and welding pressure, to improve welding quality and efficiency;

[0050] In post-welding quality inspection, weld quality inspection: After welding is completed, the 3D vision system can perform quality inspection on the weld. By comparing the three-dimensional point cloud data before and after welding, key indicators such as the weld geometry, dimensional accuracy and surface quality can be evaluated; Defect detection and identification: The system can automatically identify defects in the weld, such as cracks, pores, slag inclusions, etc., to provide data support for subsequent repairs and improvements.

[0051] like Figure 2 , 3As shown in , 4, 5, 6, 7, and 8, in steps 1 to 3, the weld extraction technology based on deep learning (Faster-RCNN) introduces the region proposal network (RPN) and candidate region pooling (RoIPooling) layer into the Faster-RCNN classic target detection deep neural network in view of the complex shape, scale changes and different levels of features of the weld image;

[0052] The residual neural network (ResNet) is used in the candidate region pooling (RoIPooling) layer. ResNet uses residual learning to connect the deep feature map with the previous shallow feature map, effectively utilizing high-level and low-level features to adapt to the detection of small targets.

[0053] The scale changes and features at different levels in weld images are taken into consideration. A multi-scale feature fusion mechanism, pyramid network (FPN), is used to fuse feature maps from different levels by introducing side connections and upsampling operations in the feature extraction network to generate a feature pyramid with multi-scale information.

[0054] In the region proposal network (RPN), according to the characteristics of weld images, the improved model adjusts the size and scale of the anchor box, adopts different sliding window strategies, or introduces a more effective candidate region generation mechanism to improve the accuracy and recall rate of the candidate region;

[0055] Classification loss function and bounding box regression loss for weld targets, weighted, balanced or multi-task learning strategies for weld features to improve the performance and robustness of the model;

[0056] The classification loss function is as follows:

[0057]

[0058] in,

[0059]

[0060] Deep learning (Faster-RCNN) weld extraction technology also includes transfer learning and model fine-tuning;

[0061] Transfer learning and model fine-tuning Transfer learning is performed based on the pre-trained model, and the model is fine-tuned for the weld feature extraction task; by fine-tuning on the weld dataset, the model can better adapt to the characteristics of the weld image and improve the detection performance and generalization ability;

[0062] During transfer learning, the following transfer loss function based on optimal transmission is used for model optimization; min π∈∏(μ,ν) ∫ x×y c(x,y)dπ(x,y).

[0063] It should be noted that, in this embodiment, in the weld feature extraction technology based on the improved Faster-RCNN, a convolutional neural network (CNN) for target detection is adopted in the design of the feature extraction network and the multi-scale feature fusion mechanism according to the characteristics of the weld image, and the residual neural network (ResNet) is specifically selected, because ResNet has a deeper convolution layer; due to the downsampling effect in the traditional convolutional neural network, small objects cannot obtain significant features; in addition, the improved model introduces a multi-scale feature fusion mechanism pyramid network (FPN), which can effectively improve the detection performance of the model for small-sized targets and has a lower computational cost;

[0064] In the improved region proposal network and weld-specific loss function design, the classification loss function and the migration loss function are used to optimize the model.

[0065] Through the above algorithm design, a comparison between the actual and predicted images of weld identification and detection was carried out on real welds. The application of this technology has achieved an accuracy of 99% in the identification of different types of weld defects. Compared with the original Faster R-CNN model, the accuracy rate has increased by 8%, the precision rate has increased by 5%, and the recall rate has increased by 7%, thus improving the accuracy of weld detection in the industry.

[0066] like Figure 1 , 9 As shown, the 3D vision in steps 1 to 3 is based on the 3D vision measurement technology based on digital micromirror technology (DLP) DLP projection. According to the parameter information of the structured light system equipment (such as resolution and grayscale range), appropriate coefficients are set to design a sinusoidal variation function to generate a grayscale stripe coding sequence; then the obtained grayscale stripe coding sequence is assigned to each row of the coding pattern, that is, the grayscale coding information of each row of the coding pattern is the same; finally, the stripes are colored with one color to obtain a monochrome sinusoidal stripe coding pattern;

[0067] All stripes in the coding pattern are coded in red, and there are black stripes between the red stripes. In addition to the different stripe color information, there is also a difference in the coding of grayscale values ​​in the horizontal direction between the proposed coding pattern and the color stripe coding pattern.

[0068] The color stripe coding pattern uses binary coding, while the proposed coding pattern uses multi-value coding and satisfies sinusoidal variation;

[0069] Based on the monochrome sinusoidal stripes, all stripes are encoded with the same color, and random points are embedded around the stripes. The random points are used to identify the stripes according to the distribution of random points around the stripes. Finally, the generated pseudo-random point pattern and the proposed monochrome sinusoidal stripe coding pattern are combined to provide coding information for the identification of monochrome stripes.

[0070] In order to make the random dots highly distinguishable in the encoded structured light pattern and the captured structured light image, regular restrictions are imposed on the random dot size, random dot density, and random dot distribution;

[0071] The coding pattern only requires two colors, which can be reconstructed and transformed according to the colored objects.

[0072] It should be noted that in this embodiment, the coding pattern designed by this technology in the 3D visual measurement technology based on DLP projection only needs two colors, the coding is simple, not easily affected by color crosstalk, and can be used to reconstruct colored objects; the two colors used are changeable, so the coding color can be changed according to the color of the measured object, which reduces the problem that the current 3D structured light camera cannot measure the reflection of the welding object or the light-colored object surface due to the single color of the structured light. At the same time, due to the introduction of random points, the coding pattern is spatially coded, and only a single projection is required, which improves the efficiency of weld measurement.

[0073] The working principle of the present invention is:

[0074] In the preparation stage before welding, weld identification and positioning: before welding begins, the 3D vision system scans the weld with a high-precision 3D camera to obtain the three-dimensional point cloud data of the weld; using intelligent algorithms, the system can extract the accurate position information of the weld, including the geometric shape, size and position coordinates of the weld; workpiece alignment and positioning: in addition to weld identification, the 3D vision system can also perform overall alignment and positioning of the workpiece to ensure the position accuracy of the workpiece before welding and reduce welding quality problems caused by workpiece position deviation;

[0075] Real-time weld tracking during welding: During welding, the 3D vision system can monitor the position and status of the weld in real time and dynamically adjust the welding trajectory to ensure that the welding gun or welding robot can accurately track the weld for welding; Welding posture adjustment: Based on the real-time weld information, the system can calculate the optimal welding posture of the welding robot or welding equipment, including parameters such as welding angle, welding speed and welding pressure, to improve welding quality and efficiency;

[0076] In the post-welding quality inspection, weld quality inspection: After welding is completed, the 3D vision system can perform quality inspection on the weld. By comparing the 3D point cloud data before and after welding, key indicators such as the weld geometry, dimensional accuracy and surface quality can be evaluated; Defect detection and identification: The system can automatically identify defects in the weld, such as cracks, pores, slag inclusions, etc., to provide data support for subsequent repairs and improvements;

[0077] In the weld feature extraction technology based on the improved Faster-RCNN, a convolutional neural network (CNN) for target detection is used in the design of the feature extraction network and the multi-scale feature fusion mechanism according to the characteristics of the weld image. Specifically, the residual neural network (ResNet) is selected because ResNet has a deeper convolution layer; due to the downsampling effect in the traditional convolutional neural network, small objects cannot obtain significant features; in addition, the improved model introduces a multi-scale feature fusion mechanism pyramid network (FPN), which can effectively improve the model's detection performance for small-sized targets and has a lower computational cost such as Figure 2 As shown;

[0078] In the improved region proposal network and weld-specific loss function design, the classification loss function and the migration loss function are used to optimize the model.

[0079] Through the above algorithm design, the real image and predicted image of weld seam recognition detection were compared on the real weld seam. The results are as follows: Figure 3-8 As shown in the figure, the application of this technology has achieved an accuracy of 99% in the recognition of different types of weld defects. Compared with the original Faster R-CNN model, the accuracy rate has increased by 8%, the precision rate has increased by 5%, and the recall rate has increased by 7%, which has improved the accuracy of weld detection in the industry.

[0080] In addition, the coding pattern designed by this technology in the 3D visual measurement technology based on DLP projection only needs two colors, the coding is simple, not easily affected by color crosstalk, and can be used to reconstruct colored objects; the two colors used are changeable, so the coding color can be changed according to the color of the measured object, which reduces the problem that the current 3D structured light camera cannot measure the reflection of the welding object or the surface of the light-colored object due to the single color of the structured light. At the same time, due to the introduction of random points, the coding pattern is spatially coded, and only a single projection is required, which improves the efficiency of weld measurement.

[0081] The above embodiments are only for illustrating the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A 3D vision welding measurement method based on deep learning, characterized in that ,The visual measurement welding method comprises the following steps: Step 1: Image measurement preparation: Before welding begins, the 3D vision system scans the weld with a high-precision 3D camera to obtain the 3D point cloud data of the weld. Using intelligent algorithms, the system can extract the accurate position information of the weld, including the geometric shape, size and position coordinates of the weld; Step 2: Image preprocessing: During the welding process, the 3D vision system can monitor the position and status of the weld in real time and dynamically adjust the welding trajectory to ensure that the welding gun or welding robot can accurately track the weld for welding; Step 3: Inspection and evaluation: After welding is completed, the 3D vision system can perform quality inspection on the weld. By comparing the 3D point cloud data before and after welding, the key indicators of the weld geometry, dimensional accuracy and surface quality can be evaluated.

2. The 3D vision measurement welding method based on deep learning according to claim 1 is characterized in that: In steps 1 to 3, based on the weld extraction technology of deep learning (Faster-RCNN), the region proposal network (RPN) and candidate region pooling (RoIPooling) layer are introduced into the Faster-RCNN classic target detection deep neural network in view of the complex shape, scale changes and features of different levels of the weld image.

3. The 3D vision measurement welding method based on deep learning according to claim 2 is characterized in that: The candidate region pooling (RoIPooling) layer uses a residual neural network (ResNet), which uses residual learning to connect deep feature maps with previous shallow feature maps, effectively utilizing high-level and low-level features to adapt to the detection of small targets; The scale changes and features at different levels in weld images are taken into account. A multi-scale feature fusion mechanism, FPN, is used. By introducing side connections and upsampling operations in the feature extraction network, feature maps from different levels are fused to generate a feature pyramid with multi-scale information.

4. The 3D vision measurement welding method based on deep learning according to claim 2 is characterized in that: In the region proposal network (RPN), according to the characteristics of the weld image, the improved model adjusts the size and proportion of the anchor box, adopts different sliding window strategies or introduces a more effective candidate region generation mechanism to improve the accuracy and recall rate of the candidate region; Classification loss function and bounding box regression loss for weld targets, weighted, balanced or multi-task learning strategies for weld features to improve the performance and robustness of the model; The classification loss function is as follows: in, 5. The 3D vision measurement welding method based on deep learning according to claim 2, characterized in that: The deep learning (Faster-RCNN) weld extraction technology also includes transfer learning and model fine-tuning; The transfer learning and model fine-tuning perform transfer learning based on a pre-trained model, and perform model fine-tuning for a weld feature extraction task; by performing fine-tuning on a weld dataset, the model can better adapt to the characteristics of weld images, thereby improving detection performance and generalization capabilities.

6. The 3D vision measurement welding method based on deep learning according to claim 5 is characterized in that: During the transfer learning, the following transfer loss function based on optimal transmission is used to optimize the model; min π∈∏(μ,ν) ∫ x×y c(x,y)dπ(x,y)。 7. The 3D vision measurement welding method based on deep learning according to claim 1, characterized in that: The 3D vision in steps 1 to 3 is based on a 3D vision measurement technology based on digital micromirror technology (DLP) DLP projection. According to the parameter information of the structured light system equipment (such as resolution and grayscale range), appropriate coefficients are set to design a sinusoidal variation function to generate a grayscale stripe coding sequence; then the obtained grayscale stripe coding sequence is assigned to each row of the coding pattern, that is, the grayscale coding information of each row of the coding pattern is the same; finally, the stripes are colored with a color to obtain a monochrome sinusoidal stripe coding pattern; All stripes in the coding pattern are coded using red, and there are black stripes between the red stripes; In addition to the different stripe color information, there is also a difference in the encoding of grayscale values ​​in the horizontal direction between the proposed coding pattern and the color stripe coding pattern; The color stripe coding pattern uses binary coding, and the proposed coding pattern uses multi-value coding and satisfies sinusoidal variation.

8. The 3D vision measurement welding method based on deep learning according to claim 7 is characterized in that: Based on the monochrome sinusoidal stripes, all stripes are encoded using the same color, and random points are embedded around the stripes. The random points are used to identify the stripes based on the distribution of random points around the stripes. Finally, the generated pseudo-random dot pattern and the proposed monochrome sinusoidal stripe coding pattern are combined to provide coding information for the identification of monochrome stripes.

9. The 3D vision measurement welding method based on deep learning according to claim 8, characterized in that: In order to make the random dots highly distinguishable in the encoded structured light pattern and the captured structured light image, regular restrictions are imposed on the random dot size, random dot density and random dot distribution.

10. The 3D vision measurement welding method based on deep learning according to claim 7, characterized in that: The coding pattern only requires two colors, and the two colors can be reconstructed and transformed according to the colored object.