SSDCNet network-based weld pipeline identification method and system

Abstract

The invention provides a weld pipeline identification method and system based on an SSDCNet network, and belongs to the technical field of deep learning, and the method comprises the steps: collecting the point cloud data of a weld pipeline, and obtaining a training set and a test set based on the point cloud data of the weld pipeline; the method comprises the following steps: constructing an SSDCNet network by using an SSDC module of a coding and decoding framework, and dividing the SSDCNet network into a classification network and a segmentation network to obtain an improved SSDCNet network; and determining preset training parameters, training the improved SSDCNet network by using the training set, and performing performance test on the improved SSDCNet network by using the test set to obtain a weld pipeline identification model. According to the method, the SSDC module is designed, so that the local features in the point cloud data can be captured, the local features of the point cloud data can be effectively extracted, and the most advanced accuracy is obtained in tasks such as target classification, target part segmentation and semantic scene segmentation.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a weld pipeline identification method and system based on an SSDCNet network. Background Art

[0002] In the process of pipeline cutting and welding, there are many links that require manual assistance. For example, when cutting different types of pipes, the cutting machine needs manual guidance on the cutting starting point and setting parameters such as the incision type. In pipeline welding projects, manual teaching and manual selection of welding points also require manual intervention.

[0003] With the rapid development of 3D sensing technologies such as LiDAR and RGB-D cameras, the availability of 3D point cloud data has increased significantly. Point clouds, a set of 3D points representing the surface of an object, have become an important data source for various applications including autonomous driving, robotics, virtual reality, and augmented reality. At the same time, deep learning-based methods have achieved remarkable success in various computer vision tasks, including image classification, object detection, and semantic segmentation. Therefore, algorithms based on raw point clouds have become a hot topic in recent years. Raw point clouds contain rich geometric and semantic information, making it easier for algorithms to achieve scene perception.

[0004] Existing point cloud processing methods have some shortcomings when processing disordered point cloud data. First, the disorder of point clouds makes it impossible for traditional convolutional neural networks to directly process point cloud data, and they need to be converted into ordered data formats such as grids or voxels. However, this conversion may lead to the loss of geometric information, increased computational complexity, and increased data size. Secondly, existing point cloud processing methods have a certain trade-off between extracting local features and maintaining the disorder of point clouds. Some methods use maximum pooling and symmetric functions to extract features, but this may lead to the loss of local features. Other methods extract features by learning local convolution order, but their training speed is slow. In addition, existing point cloud processing methods may have limitations when processing objects of different scales and lack the ability to handle scale invariance. Summary of the invention

[0005] The present invention provides a weld pipeline identification method and system based on SSDCNet network, so as to solve the defects of point cloud processing for pipeline cutting and welding in the prior art, such as complex calculation, easy loss of local features, and limitations in processing objects of different scales.

[0006] In a first aspect, the present invention provides a weld pipeline identification method based on the SSDCNet network, comprising: Collecting weld pipeline point cloud data, and obtaining a training set and a test set based on the weld pipeline point cloud data; The SSDCNet network is constructed by using the spherical shell directional convolution SSDC module of the codec framework, and the SSDCNet network is divided into a classification network and a segmentation network to obtain an improved SSDCNet network; Determine preset training parameters, use the training set to train the improved SSDCNet network, use the test set to perform performance testing on the improved SSDCNet network, and obtain a weld pipeline recognition model; The point cloud data of the weld pipe to be identified is input into the weld pipe identification model, and the weld pipe classification and the three-dimensional segmentation point cloud of the weld pipe are output.

[0007] According to a weld pipeline identification method based on SSDCNet network provided by the present invention, weld pipeline point cloud data is collected, and a training set and a test set are obtained based on the weld pipeline point cloud data, including: The point cloud processing software couldcompare is used to mark the weld pipeline point cloud data to distinguish the pipeline area of ​​the workpiece; The annotated weld pipeline point cloud data is divided into the training set and the test set.

[0008] According to a weld pipeline identification method based on SSDCNet network provided by the present invention, the SSDCNet network is constructed by using the SSDC module of the encoding and decoding framework, including: The SSDCNet network includes three coding layers, namely the first coding layer, the second coding layer and the third coding layer, each coding layer includes an SSDC module; The number of point clouds included in the first coding layer, the second coding layer, and the third coding layer decreases in sequence, and the number of output channels included increases in sequence, and the radius of each coding layer is the number of point clouds divided by a preset coefficient; The Adam optimizer and the preset initial learning rate are used for optimization to obtain the SSDCNet network.

[0009] According to a weld pipeline identification method based on an SSDCNet network provided by the present invention, the SSDCNet network is divided into a classification network and a segmentation network to obtain an improved SSDCNet network, including: the classification network is a two-dimensional convolutional neural network, the encoder of the classification network includes three SSDC modules connected in sequence, the number of point clouds of the three SSDC modules gradually decreases, and the number of output feature channels gradually increases to obtain an output matrix, the output matrix is ​​input into a multi-layer perceptron module to generate an object classification probability map matrix, and the object classification probability map matrix includes the number of categories; The segmentation network is a U-Net network, and the encoder of the classification network includes three SSDC modules connected in sequence. The number of point clouds of the three SSDC modules gradually increases, and the number of output feature channels gradually decreases. The object classification probability map matrix is ​​input into the first SSDC module, and the point cloud recognition data is output by the third SSDC module.

[0010] According to a weld pipeline identification method based on SSDCNet network provided by the present invention, the input of the SSDC module is an n×d matrix, where n is the number of points in the point cloud, and d is the d-dimensional feature of each point; The SSDC module includes a spherical shell convolution module and a direction encoding module.

[0011] According to a weld pipeline identification method based on SSDCNet network provided by the present invention, the spherical shell convolution module specifically includes: Determine any point in the spherical shell convolution module, obtain a set of neighboring points of the any point, and convolve all neighboring point input features in the neighboring point set with neighboring point convolution weights to obtain an input feature of the any point; Based on the Euclidean distance between any point and all neighboring points in the neighboring point set, a plurality of spherical shells are obtained, and a spherical shell convolution feature is obtained using the plurality of spherical shells and the weight of each spherical shell.

[0012] According to a weld pipeline identification method based on SSDCNet network provided by the present invention, the direction encoding module specifically includes: The spherical shell layers corresponding to the multiple spherical shells are divided into 8 regions along the 8 directions of the coordinate axis, and a neighboring search is performed in each region to screen the neighboring point closest to any point. If no neighboring point is found in any region, the any point is used as the neighboring point; A direction-encoded convolution operator is added to the eight regions, and the direction-encoded convolution operator is used to convolve a cube of a specified size along the X-axis, the Y-axis, and the Z-axis in sequence to obtain a feature vector of an adjacent point.

[0013] According to a weld pipeline identification method based on the SSDCNet network provided by the present invention, preset training parameters are determined, the improved SSDCNet network is trained using the training set, and the improved SSDCNet network is performance tested using the test set to obtain a weld pipeline identification model, including: Determine the learning rate, learning rate momentum, batch size, total training rounds, weight decay and maximum number of iterations, and input the training set into the improved SSDCNet network for training; The test set is input into the trained model for testing to obtain the weld pipeline recognition model.

[0014] In a second aspect, the present invention further provides a weld pipeline identification system based on the SSDCNet network, comprising: An acquisition module, used for acquiring weld pipeline point cloud data, and obtaining a training set and a test set based on the weld pipeline point cloud data; A construction module, used to construct an SSDCNet network using the SSDC module of the codec framework, and divide the SSDCNet network into a classification network and a segmentation network to obtain an improved SSDCNet network; A training module, used to determine preset training parameters, train the improved SSDCNet network using the training set, perform performance testing on the improved SSDCNet network using the test set, and obtain a weld pipeline recognition model; The identification module is used to input the point cloud data of the weld pipe to be identified into the weld pipe identification model, and output the weld pipe classification and the three-dimensional segmentation point cloud of the weld pipe.

[0015] In a third aspect, the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a weld pipe identification method based on the SSDCNet network as described above is implemented.

[0016] In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described weld pipe identification methods based on the SSDCNet network.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The weld pipeline identification method and system based on the SSDCNet network provided by the present invention can capture local features in point cloud data by designing the SSDC module, divide the neighborhood set of sampling points into different spherical shell layers, extract features along different coordinate axes through spherical shell convolution, and perform convolution operations in sequence from the inside to the outside, which reflects the scale invariance of the network, can effectively extract local features of point cloud data, and achieve state-of-the-art accuracy in tasks such as target classification, target part segmentation, and semantic scene segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 It is a flow chart of the weld pipeline identification method based on the SSDCNet network provided by the present invention; Figure 2 It is the SSDCNet network structure provided by the present invention; Figure 3 It is a local structure diagram of the SSDC spherical operator provided by the present invention; Figure 4 It is a front view and a side view of the SSDC spherical operator provided by the present invention; Figure 5 It is a structural schematic diagram of a weld pipeline identification system based on the SSDCNet network provided by the present invention; Figure 6 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are 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.

[0021] Existing point cloud processing methods have some shortcomings when processing disordered point cloud data. First, the disorder of point clouds makes it impossible for traditional convolutional neural networks to directly process point cloud data, and they need to be converted into ordered data formats such as grids or voxels. However, this conversion may lead to the loss of geometric information, increased computational complexity, and increased data size. Secondly, existing point cloud processing methods have a certain trade-off between extracting local features and maintaining the disorder of point clouds. Some methods use maximum pooling and symmetric functions to extract features, but this may lead to the loss of local features. Other methods extract features by learning local convolution order, but their training speed is slow. In addition, existing point cloud processing methods may have limitations when processing objects of different scales and lack the ability to handle scale invariance.

[0022] In view of the problems existing in the prior art, the present invention proposes a method for object recognition and segmentation based on the spherical surface direction convolution operator (SSDCNet) network, which has the following characteristics: (1) it focuses on the structured properties of point clouds, considers the complex local structure of sampling points, and obtains local feature information along directions and levels; (2) it has strong local feature expression capabilities and a high recognition rate for edges, textures, and corners.

[0023] Figure 1 It is a schematic flow chart of the weld pipe recognition method based on the SSDCNet network provided by an embodiment of the present invention. As Figure 1 shown, it includes: Step 100: Collect the weld pipe point cloud data, and obtain a training set and a test set based on the weld pipe point cloud data; Step 200: Use the spherical shell direction convolution SSDC module of the encoding-decoding framework to construct the SSDCNet network, divide the SSDCNet network into a classification network and a segmentation network, and obtain an improved SSDCNet network; Step 300: Determine the preset training parameters, use the training set to train the improved SSDCNet network, use the test set to perform performance testing on the improved SSDCNet network, and obtain a weld pipe recognition model; Step 400: Input the weld pipe point cloud data to be recognized into the weld pipe recognition model, and output the weld pipe classification and the three-dimensional segmentation point cloud of the weld pipe.

[0024] Specifically, in the embodiment of the present invention, the weld pipe point cloud data is first collected, and the point cloud annotation software couldcompare is used to annotate and produce a data set, and the specific workpiece pipe area is annotated and divided into a training set and a test set.

[0025] Then, an encoding-decoding (downsampling-upsampling) framework is adopted, and the SSDCNet network is constructed with the SSDC module. The SSDCNet network is divided into two types: a classification network and a segmentation network. The classification network is used to classify the pipe part of the original point cloud, and the segmentation network is used to segment the pipe part in the original point cloud.

[0026] Among them, in the classification network, three SSDC modules are connected in sequence as the encoder part. The points of the input point cloud are gradually subsampled. The number of representative points decreases layer by layer, expressed as N0>N1>N2, (N represents the number of points in the input point cloud), and the number of output feature channels increases layer by layer, expressed as C0<C1<C2, (C represents the number of feature channels). The segmentation network follows the U-Net network structure, which is an encoder-decoder architecture with skip connections. The deconvolution part starts from the N×C set obtained from the last layer of the encoder, and through the SSDC operator until the point cloud reaches the original resolution. The deconvolution layer gradually generates more points but fewer feature channels. The skip connections retain the features of the earlier layers and connect them with the output features of the deconvolution layer.

[0027] Next, the improved network is trained using the collected training set; the learning rate, learning rate momentum, batch size, total training rounds, weight decay and maximum number of iterations are set as training parameters to train the improved SSDCNet network. The test data set is input into the SSDCNet network for testing to obtain the weld pipe recognition model.

[0028] Finally, in practical applications, the point cloud data of the weld pipe to be identified is input into the trained weld pipe recognition model to obtain the classified weld pipe and the segmented weld pipe three-dimensional point cloud.

[0029] By designing the SSDC module, the present invention can capture local features in point cloud data, divide the neighborhood set of sampling points into different spherical shell layers, extract features along different coordinate axes through spherical shell convolution, and perform convolution operations sequentially from the inside to the outside, which reflects the scale invariance of the network. It can effectively extract local features of point cloud data and achieve state-of-the-art accuracy in tasks such as target classification, target part segmentation, and semantic scene segmentation.

[0030] In one embodiment, since the spatial distribution characteristic of the point cloud P is its irregularity, the density in different directions is different. In addition, the point cloud P lacks a recognizable arrangement or framework, which is different from traditional data formats such as images or videos, which are usually organized in a sequential or structured manner. In order to achieve point cloud recognition and segmentation tasks, an embodiment of the present invention develops a deep neural network that specifically processes point clouds as input. The point cloud deep learning network is named SSDCNet and consists of three layers. SSDCNet adopts an encoding-decoding (downsampling-upsampling) framework with three encoding layers. Each encoding layer contains an SSDC module. The parameters Ni, di, and Ci represent the number of representative points, the radius of each layer, and the number of output channels, respectively. In an embodiment of the present invention, from the first layer to the third layer, for i = 0, 1, 2, Ni is set to 512, 128, and 32, di is set to the point cloud density divided by 50, and Ci is set to 128, 256, and 512. In the last layer of convolution for segmentation, C is set to 64. During the training process, 32 samples are used for classification training and 16 samples are used for segmentation training. The optimization process uses the Adam optimizer with an initial learning rate of 0.001.

[0031] The SSDCNet network is divided into two types: classification network and segmentation network. The classification network is used to classify the pipeline part of the original point cloud. The segmentation network is used to segment the pipeline part in the original point cloud. Specifically, Figure 2As shown in the figure, in the classification network, three SSDC modules are connected in sequence as the encoder part. The points of the input point cloud are gradually subsampled, and the number of representative points decreases layer by layer, denoted as N0 > N1 > N2, (N represents the number of points in the input point cloud), while the number of output feature channels increases layer by layer, denoted as C0 < C1 < C2 (C represents the number of feature channels). This design is similar to a typical 2D convolutional neural network: the number of representative points decreases while the number of output channels increases. After three layers of SSDC, a matrix of size N2×C2 is obtained, where N2 is the number of final representative points extracted from the input point cloud, and each point contains a high-dimensional feature vector of size C2. This matrix is input into a Multilayer Perceptron (MLP) module to generate a probability map for object classification. Finally, a matrix of size 128×k is obtained, where k represents the number of classes.

[0032] The segmentation network follows the U-Net network structure, which is an encoder-decoder architecture with skip connections. The deconvolution part starts from the N×C set obtained from the last layer of the encoder and passes through the SSDC operator until the point cloud reaches the original resolution. The deconvolution layer gradually generates more points but fewer feature channels. The skip connections retain the features of the earlier layers and connect them to the output features of the deconvolution layer. This strategy can effectively improve the recognition accuracy in the dense semantic segmentation of point clouds.

[0033] It should be noted that in the network input, the point cloud is defined as a set of three-dimensional points {Pi | i = 1,..., n}, where each point Pi contains a vector representing its (x, y, z) coordinates, as well as other feature channels such as color, normal, etc. For simplicity and clarity, unless otherwise specified, only the (x, y, z) coordinates are used as the channels of the points.

[0034] When the point cloud is considered as a whole dataset, it is inherently disordered and lacks a fixed framework and arrangement. However, for a single point P, the neighboring points K around it have directionality and order in the spatial structure with respect to point p. At the same time, there are also different spatial relationships among different K points. Therefore, when obtaining the local features of point p, the direction and distance of the features should also be considered.

[0035] The input of the SSDC module is an n×d matrix, where n is the number of points in the point cloud and d represents the d-dimensional features (x, y, z coordinates) of each point, and the output is an n×d matrix, with each point updated with new d-dimensional features. The module consists of two parts: spherical shell convolution and direction encoding. The design and principle of each part will be introduced in detail in the following paragraphs.

[0036] In the spherical shell convolution module, for the input point set, a set of representative points is obtained by random sampling. For each representative point p, the surrounding neighboring points are obtained, where Ωp is the neighbor set determined by the nearest neighbor query. Therefore, the convolution at each representative point p is:

[0037] In the formula, F represents the input features of a channel point set; represents the convolution weight of each point. F(p) and F(q) represent the features of points p and q.

[0038] In this formula, Ωp can be simply regarded as a mathematical point set. For each q point in the Ωp set, they are discrete and have no correlation with each other, which further reflects the disorder of the point cloud set. Therefore, the w weight in this formula is discretized, and a w weight needs to be defined for each point separately. This is impractical in neural network training. In order to solve this problem, the q points need to be sorted or divided into different areas.

[0039] By observation, we can know that although each q point in the set is unrelated, for point p, each q point in the Ωp set has a different Euclidean distance with point p. In the neighborhood point set of point p, there is a direction from inside to outside.

[0040] Therefore, the neighbor set Ωp of each representative point is divided into different layers according to the Euclidean distance from point p. Each layer contains a different number of neighboring points, which are stacked layer by layer to form a spherical shell structure. Each layer of the spherical shell is given a weight w, so that w can be well defined and the output can be calculated efficiently.

[0041] In order to facilitate the neighborhood search, concentric spheres with different radii are defined. A spherical shell region S is formed between two spheres with different radii. Since the spherical shells are naturally ordered (from the innermost to the outermost), there is an order between the spherical shells. Figure 3 is the local structure diagram of SSDC spherical operator, Figure 4 is the front view and side view of the SSDC operator. Therefore, the convolution can be well defined, and the weight of each spherical shell is . The convolution can be defined as:

[0042] In the formula Represents the weight of each spherical shell. i represents the spherical shell layer from the innermost to the outermost (three layers in total). Represents the characteristics of the spherical shell layer S.

[0043] Previous local layer feature methods usually use unordered operations such as max pooling. This is based on the understanding that point clouds are unordered and unstructured. However, using ordered operators can provide more information compared to unordered operations such as max pooling, which discards all inputs except the maximum value. In order to maintain the invariance of the order of the input points, it is important to find a way to represent the directionality of the points. For each spherical shell layer S, the points in it are still unordered. But each point itself is directional in terms of spatial properties, which is determined by the three coordinate features of the point itself. Therefore, a simplified orientation encoding (OE) unit is used in the orientation encoding module to obtain the features of each layer of the spherical shell. The OE unit takes the d-dimensional feature vector f0∈Rd of point p0 as input. A two-stage scheme is adopted to integrate the information from eight directions and generate a direction-aware feature f.

[0044] In the first stage of the OE unit, the spherical shell layer is divided into 8 regions L along the 8 directions of the coordinate axis. In each region, a neighbor search is used to select the nearest neighboring point to the sampling point. If there is no point in a region, the representative point p is used as the nearest neighboring point in the region.

[0045] In the second stage, the features of neighboring points in 8 regions are processed. Many previous works have ignored the importance of considering the structure of the data. They usually apply max pooling to the feature vector along the d dimension to generate new features. However, it is believed that ordered operators, such as convolution, have the potential to better exploit the inherent structure of the data. Therefore, in this study, a new operator called directional encoding convolution is proposed. This three-stage operator convolves a 2×2×2 cube along the X, Y, and Z axes in sequence. Formally, the features of neighboring points are represented by a vector L of shape 2×2×2×d, where the first three dimensions correspond to the three axes. Each slice of the vector L represents a feature vector, where L(1,1,1) represents the features from the top front right eighth. The directional convolution can be expressed as follows:

[0046]

[0047]

[0048] in , , are the weights of the convolution operator, and g(·) = ReLU(·) represents the activation function. Represents the combination of 8 spatial direction features in the spherical shell layer S.

[0049] In one embodiment, the collected training set is used to train the improved network; the learning rate, learning rate momentum, batch size, total training rounds, weight decay and maximum number of iterations are set as training parameters to train the improved SSDCNet network; and the test data set is input into the SSDCNet network for testing to obtain a weld pipeline recognition model.

[0050] The weld pipe identification system based on the SSDCNet network provided by the present invention is described below. The weld pipe identification system based on the SSDCNet network described below and the weld pipe identification method based on the SSDCNet network described above can be referenced to each other.

[0051] Figure 5 is a structural diagram of a weld pipeline identification system based on SSDCNet network provided by an embodiment of the present invention, such as Figure 5 As shown, it includes: a collection module 51, a construction module 52, a training module 53 and a recognition module 54, wherein: The acquisition module 51 is used to acquire weld pipeline point cloud data, and obtain a training set and a test set based on the weld pipeline point cloud data; the construction module 52 is used to use the SSDC module of the codec framework to construct the SSDCNet network, divide the SSDCNet network into a classification network and a segmentation network, and obtain an improved SSDCNet network; the training module 53 is used to determine preset training parameters, use the training set to train the improved SSDCNet network, and use the test set to perform a performance test on the improved SSDCNet network to obtain a weld pipeline recognition model; the recognition module 54 is used to input the weld pipeline point cloud data to be identified into the weld pipeline recognition model, and output the weld pipeline classification and the weld pipeline three-dimensional segmentation point cloud.

[0052] Figure 6 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 6As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630 and a communication bus 640, wherein the processor 610, the communication interface 620 and the memory 630 communicate with each other through the communication bus 640. The processor 610 may call the logic instructions in the memory 630 to execute the weld pipe identification method based on the SSDCNet network, the method comprising: collecting weld pipe point cloud data, obtaining a training set and a test set based on the weld pipe point cloud data; constructing an SSDCNet network using the spherical shell directional convolution SSDC module of the codec framework, dividing the SSDCNet network into a classification network and a segmentation network, and obtaining an improved SSDCNet network; determining preset training parameters, training the improved SSDCNet network using the training set, and performing a performance test on the improved SSDCNet network using the test set to obtain a weld pipe identification model; inputting the weld pipe point cloud data to be identified into the weld pipe identification model, and outputting weld pipe classification and weld pipe three-dimensional segmentation point cloud.

[0053] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0054] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0055] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A weld pipeline identification method based on SSDCNet network, characterized in that: include: Collecting weld pipeline point cloud data, and obtaining a training set and a test set based on the weld pipeline point cloud data; The SSDCNet network is constructed by using the spherical shell directional convolution SSDC module of the codec framework, and the SSDCNet network is divided into a classification network and a segmentation network to obtain an improved SSDCNet network; Determine preset training parameters, use the training set to train the improved SSDCNet network, use the test set to perform performance testing on the improved SSDCNet network, and obtain a weld pipeline recognition model; The point cloud data of the weld pipe to be identified is input into the weld pipe identification model, and the weld pipe classification and the three-dimensional segmentation point cloud of the weld pipe are output.

2. The weld pipeline identification method based on SSDCNet network according to claim 1 is characterized in that: Collecting weld pipeline point cloud data, and obtaining a training set and a test set based on the weld pipeline point cloud data, including: The point cloud processing software couldcompare is used to mark the weld pipeline point cloud data to distinguish the pipeline area of ​​the workpiece; The annotated weld pipeline point cloud data is divided into the training set and the test set.

3. The weld pipeline identification method based on SSDCNet network according to claim 1 is characterized in that: The SSDCNet network is constructed using the SSDC module of the codec framework, including: The SSDCNet network includes three coding layers, namely the first coding layer, the second coding layer and the third coding layer, each coding layer includes an SSDC module; The number of point clouds included in the first coding layer, the second coding layer, and the third coding layer decreases in sequence, and the number of output channels included increases in sequence, and the radius of each coding layer is the number of point clouds divided by a preset coefficient; The Adam optimizer and the preset initial learning rate are used for optimization to obtain the SSDCNet network.

4. The weld pipeline identification method based on SSDCNet network according to claim 3 is characterized in that: The SSDCNet network is divided into a classification network and a segmentation network to obtain an improved SSDCNet network, including: the classification network is a two-dimensional convolutional neural network, the encoder of the classification network includes three SSDC modules connected in sequence, the number of point clouds of the three SSDC modules gradually decreases, and the number of output feature channels gradually increases to obtain an output matrix, the output matrix is ​​input into a multi-layer perceptron module to generate an object classification probability map matrix, and the object classification probability map matrix includes the number of categories; The segmentation network is a U-Net network, and the encoder of the classification network includes three SSDC modules connected in sequence. The number of point clouds of the three SSDC modules gradually increases, and the number of output feature channels gradually decreases. The object classification probability map matrix is ​​input into the first SSDC module, and the point cloud recognition data is output by the third SSDC module.

5. The weld pipeline identification method based on SSDCNet network according to claim 3 is characterized in that: The input of the SSDC module is an n×d matrix, where n is the number of points in the point cloud and d is the d-dimensional feature of each point; The SSDC module includes a spherical shell convolution module and a direction encoding module.

6. The weld pipeline identification method based on SSDCNet network according to claim 5 is characterized in that: The spherical shell convolution module specifically includes: Determine any point in the spherical shell convolution module, obtain a set of neighboring points of the any point, and convolve all neighboring point input features in the neighboring point set with neighboring point convolution weights to obtain an input feature of the any point; Based on the Euclidean distance between any point and all neighboring points in the neighboring point set, a plurality of spherical shells are obtained, and a spherical shell convolution feature is obtained using the plurality of spherical shells and the weight of each spherical shell.

7. The weld pipeline identification method based on SSDCNet network according to claim 6 is characterized in that: The direction encoding module specifically includes: The spherical shell layers corresponding to the multiple spherical shells are divided into 8 regions along the 8 directions of the coordinate axis, and a neighboring search is performed in each region to screen the neighboring point closest to any point. If no neighboring point is found in any region, the any point is used as the neighboring point; A direction-encoded convolution operator is added to the eight regions, and the direction-encoded convolution operator is used to convolve a cube of a specified size along the X-axis, the Y-axis, and the Z-axis in sequence to obtain a feature vector of an adjacent point.

8. The weld pipeline identification method based on SSDCNet network according to claim 1 is characterized in that: Determine preset training parameters, use the training set to train the improved SSDCNet network, use the test set to perform performance testing on the improved SSDCNet network, and obtain a weld pipeline recognition model, including: Determine the learning rate, learning rate momentum, batch size, total training rounds, weight decay and maximum number of iterations, and input the training set into the improved SSDCNet network for training; The test set is input into the trained model for testing to obtain the weld pipeline recognition model.

9. A weld pipeline identification system based on SSDCNet network, characterized in that: include: An acquisition module, used for acquiring weld pipeline point cloud data, and obtaining a training set and a test set based on the weld pipeline point cloud data; A construction module, used to construct an SSDCNet network using the SSDC module of the codec framework, and divide the SSDCNet network into a classification network and a segmentation network to obtain an improved SSDCNet network; A training module, used to determine preset training parameters, train the improved SSDCNet network using the training set, perform performance testing on the improved SSDCNet network using the test set, and obtain a weld pipeline recognition model; The identification module is used to input the point cloud data of the weld pipe to be identified into the weld pipe identification model, and output the weld pipe classification and the three-dimensional segmentation point cloud of the weld pipe.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, it implements the weld pipeline identification method based on the SSDCNet network as described in any one of claims 1 to 8.

Images