A workpiece point cloud data completion method based on SGAN network

Abstract

The present invention proposes a method for completing workpiece point cloud data based on an SGAN network. The method completes the point cloud in two stages: in the first stage, feature extraction is performed through an encoder based on a channel space fusion attention mechanism, and then a style-based folding network is designed as a decoder to form a preliminary coarse point cloud; in the second stage, the final point cloud is obtained through a point cloud refinement module, and then a differentiable point cloud discriminator module is used as a discriminator to perform adversarial training with the previous decoder. Through adversarial training, an image-based discriminator is used to perform adversarial improvement on the quality of the point cloud, thereby obtaining a complete point cloud. The present invention adopts a two-stage completion method based on an SGAN network to complete the missing details while ensuring that the global workpiece point cloud data is not lost.

Description

Technical Field

[0001] The present invention belongs to the technical field of joint deep learning and 3D point cloud completion, and specifically is a workpiece point cloud data completion method based on SGAN network. Background Art

[0002] With the rapid development of 3D technology, point clouds are increasingly used in the fields of autonomous driving, augmented reality, robotics, and industrial manufacturing. In addition, the acquisition of point clouds is becoming more and more convenient due to the increasing use of binocular cameras, structured light cameras, laser scanning cameras, and other devices. However, the acquired point clouds are usually noisy or incomplete, which reduces the accuracy of target recognition and segmentation based on point clouds. For example, in an industrial production environment, workpieces from workshops are reflective of metal materials and are blocked by their own structures. Using a 3D camera to directly acquire the complete original point cloud of the workpiece results in the loss of some information in the point cloud, making it difficult to perform subsequent tasks based on the workpiece point cloud. Therefore, point cloud completion technology involves generating an estimated complete point cloud from a partial point cloud, and plays a vital role in 3D point cloud tasks.

[0003] Industrial production is closely related to our lives. In the face of the requirements of industrial production modernization, many related technical applications have emerged, and point cloud is one of them. For workpieces in industrial production, the cost of obtaining a point cloud with high completion and complete three-dimensional information is very high, but deep learning technology provides convenience for point cloud completion. The point cloud completion network should meet the following three requirements at the same time: first, the output point cloud must completely retain the detailed structure of the input residual point cloud; second, the network has strong reasoning ability and can infer the overall shape from partial point cloud clues; third, the local structure should be clear, accurate, and not interfered by noise points. This requires us to use deep learning technology to highlight details while ensuring the overall characteristics.

[0004] Point cloud completion technology currently has certain demands in the field of industrial applications. With the advancement of technology, two-dimensional vision can no longer meet the production needs of industry, and research and application in the three-dimensional field have been carried out. However, due to the limitations of technical conditions such as equipment, the point cloud data of objects obtained is not complete and detailed. Many objects cannot obtain comprehensive three-dimensional data due to their own characteristics (such as reflection, occlusion), because the obtained point cloud is not comprehensive, it is impossible to perform subsequent processing such as segmentation, recognition, and measurement on the point cloud.

[0005] The point cloud of the workpiece in the factory is completed. Because the workpiece itself is small in size and has reflective and complex structures, the original point cloud obtained is incomplete, which is not conducive to the subsequent segmentation, recognition and measurement of the point cloud. Therefore, the present invention proposes an SGAN network to complete the incomplete workpiece point cloud in two stages. In the first stage, an encoder based on the channel space fusion attention mechanism is used for encoding, and then a decoder based on the style folding network is used for generation to obtain the rough point cloud data of the workpiece. In the second stage, the original input workpiece's incomplete point cloud data and the rough point cloud output in the first stage are sampled by minimum density and residual network to obtain the completed point cloud of the workpiece. Finally, a differentiable renderer and convolution operation are combined into a discriminator, which is trained against the previous decoder to improve the model. Summary of the invention

[0006] In view of the shortcomings of the existing methods, the present invention proposes a workpiece point cloud data completion method based on the SGAN network, which uses a structured light device to shoot a workpiece point cloud with incomplete information, and then uses the SGAN network to perform two-stage processing on the point cloud, and finally obtains complete workpiece point cloud information.

[0007] In a first aspect, the present invention provides a method for completing workpiece point cloud data based on an SGAN network, comprising the following steps:

[0008] Step 1: Obtain workpiece residual cloud data through a structured light camera;

[0009] Step 2: The workpiece defect cloud data is completed through the SGAN network;

[0010] The SGAN network includes an SGANNet network, a point cloud refinement module, and a differentiable point cloud discriminator module;

[0011] The SGANNet network performs the first-stage completion on the workpiece residual point cloud data to obtain the first-stage completed coarse point cloud; the SGANNet network includes an edge convolution encoder based on a channel space fusion attention mechanism and a decoder based on style folding;

[0012] The point cloud refinement module performs a second-stage completion on the coarse point cloud data completed in the first stage to obtain a second-stage completed point cloud P; the point cloud refinement module includes a minimum density sampling module and a residual network;

[0013] The differentiable point cloud discriminator module performs adversarial training on the point cloud completed in the second stage and the generator network; the differentiable point cloud discriminator module is a discriminator that combines a differentiable renderer and a 2D convolutional layer, specifically, firstly, the point cloud completed in the second stage is rendered from a 3D point cloud into a 2D depth map by the discriminator, and the probability is obtained by the 2D convolutional layer, and then the discriminator and the generator are used for adversarial training;

[0014] In a second aspect, the present invention provides a workpiece point cloud data completion system, comprising:

[0015] The data acquisition module is responsible for acquiring the workpiece residual defect cloud data;

[0016] The data completion module is responsible for completing the workpiece defect cloud data through the SGAN network.

[0017] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method described.

[0018] In a fourth aspect, the present invention provides a computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the described method is implemented.

[0019] The beneficial effects of the present invention are:

[0020] The present invention proposes a workpiece point cloud data completion method based on the SGAN network. As a whole, the innovative SGAN network is used to perform two-stage completion on the point cloud. In the first stage, an edge convolution encoder based on the channel space fusion attention mechanism and a decoder based on the style folding network are used to obtain a coarse point cloud of the workpiece. Then, in the second stage, minimum density sampling and an adversarial differentiable point cloud discriminator are used to obtain the complete completed workpiece point cloud data.

[0021] The present invention adopts a two-stage completion method based on the SGAN network to complete the missing details while ensuring that the global workpiece point cloud data is not lost. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a diagram of the SGAN network architecture of the present invention;

[0023] Figure 2 This is a diagram of the edge convolution encoder architecture based on the channel space fusion attention mechanism of the present invention;

[0024] Figure 3 The decoder architecture diagram of the present invention based on pattern folding;

[0025] Figure 4 This is the architecture diagram of the point cloud refinement module of the present invention. DETAILED DESCRIPTION

[0026] The present invention will be described below. Figure 1 The model uses the SGAN network for two-stage completion for specific explanation.

[0027] This embodiment provides a method for completing workpiece point cloud data based on an SGAN network, comprising the following steps:

[0028] Step 1: Obtain workpiece residual cloud data through a structured light camera.

[0029] Step 2: The workpiece defect cloud data is completed through the SGAN network;

[0030] 2-1 Input the workpiece residual point cloud data in step 1 into the SGANNet network to obtain the rough point cloud completed in the first stage;

[0031] The SGANNet network includes an edge convolution encoder based on a channel space fusion attention mechanism and a decoder based on style folding;

[0032] (1) Edge convolution encoder based on channel space fusion attention mechanism

[0033] The edge convolution encoder based on channel space fusion attention mechanism includes three pathways and a spatial attention module, such as Figure 2 As shown:

[0034] The first path specifically uses N workpiece residual defect cloud data P0 as encoder input, and obtains N k directed edges through a k nearest neighbor algorithm; N k directed edges pass through the first multi-layer perceptron F1 to obtain k edge features Eigenvector After an average pooling, the output is Then After passing through the second multi-layer perceptron (MLP) and sigmoid layer, the output is obtained

[0035] The second path is to and Combine and calculate the vector η, which can be expressed as:

[0036]

[0037] Among them, δ is the sigmoid function, which is used to transform each edge feature Multiply by η to recalibrate it; ° represents the mapping multiplication, F2 represents the second multi-layer perceptron (MLP) and sigmoid operation; i∈[1,N], j∈[1,k];

[0038] Finally, the vector η is activated by maximum pooling and ReLu to transform the k edge features Reduced to a new point feature P'0;

[0039] The third path is to pass the encoder input through the linear layer F3 to get N point output Then combine it with the output of the first path The new feature point P'0 obtained by the second path is added to obtain N outputs

[0040] The spatial attention module takes N outputs The channel-based global maximum pooling and global average pooling are used, and then the two pooling results are spliced ​​together on the channel. The spliced ​​results are processed by the convolution layer to reduce the channel dimension to 1, and then the sigmoid activation function is used to calculate the final feature map P out , denoted as shape code S g .

[0041] (2) Decoder based on style folding

[0042] Due to the small size and complex structure of the workpiece, some local detail features need to be controlled singly. Therefore, the present invention makes improvements on the generator based on the StyleGAN network. The encoder module based on style folding uses shape encoding and a unit square [0,1] 2 The N 2D coordinates sampled in the model are connected as input, and then pass through three style folding layers. The structure of each style folding layer is the same. The sampling features of the points are linearly transformed and batch normalized, and then linearly transformed with the shape encoding after the linear layer processing to obtain a new activation function. Finally, the new activation function is input into the attention mechanism encoder module based on channel space fusion to obtain the final output, which is the completed workpiece point cloud P. In this process, the expansion loss is used to prevent the overlap of surface elements, so that the workpiece point cloud data is evenly distributed.

[0043] Specifically, Figure 3 As shown, the decoder based on style folding encodes the shape of the encoder output S g The grid coordinates are input into the generative network at the same time; each layer of the generative network will go through 3 identical style-based folding layers, and then undergo an affine transformation through a linear layer, and then merge to obtain the final output P c , which is the coarse point cloud completed in the first stage.

[0044] The style-based folding layers are as follows:

[0045] First, the grid coordinates P in From the N-dimensional coordinate point S in It is composed of a small batch of point activations through the linear layer to obtain an N-dimensional output S, denoted as h in ;h in Through batch normalization, we get

[0046] In the same layer of the generative network, the first style-based folding layer coordinate point S in Use the unit square [0,1] 2 , the second to third style-based folding layer coordinate points S in Take the output of the previous style-based folding layer;

[0047] It can be expressed as:

[0048]

[0049] in and They are respectively in The mean and standard deviation of the activations are calculated independently for each channel.

[0050] Shape Code S g After the linear layer, we get two modulation parameters α g and β g ; Combined with the modulation parameter α g and β g Denormalize to get h out ,h out It can be expressed as:

[0051]

[0052] Finally, h out Updated to the encoder input, it is input again to the encoder based on the channel space fusion attention mechanism to obtain the output P' out .

[0053] The decoder generates h out At the same time, an expansion loss is introduced The existence of the expansion loss can ensure that the points between different surfaces do not overlap, thereby generating complex surface structures more effectively, and at the same time shrinking some sparse points to a more compact area to obtain the final output P c , that is, the coarse point cloud completed in the first stage;

[0054] Expansion loss It can be expressed as:

[0055]

[0056] Where K is the number of generated surfaces, u,v are P c Any two points in , where dis(u,v) is the Euclidean distance between points u and v is the average length of the middle side, For a filter length less than αl i The indicator function of the edge.

[0057] 2-2 Input the coarse point cloud data completed in the first stage into the point cloud refinement module to obtain the point cloud P completed in the second stage;

[0058] The segmentation of the workpiece point cloud may not require refined processing, but the measurement of the workpiece requires the refinement of the point cloud. Only with the refinement of the point cloud can the measurement accuracy be higher, especially for some workpieces that have an error range and are relatively precise, which is impossible with traditional visual technology, but can be done with three-dimensional point clouds, because the distance of the captured three-dimensional point cloud is the actual distance in the real world, so the finer the point cloud, the higher the measurement accuracy. Therefore, it is necessary to refine the workpiece point cloud. The point cloud refinement module consists of two parts: the first part is the minimum density sampling, which uses the sum of Gaussian weights to estimate the density, and then inputs the sampling result into the residual network ResNet50, and then merges it with the result obtained by the minimum density sampling to obtain the completed workpiece point cloud.

[0059] The point cloud refinement module includes a minimum density sampling module and a residual network, such as Figure 4 Specifically:

[0060] The minimum density sampling module combines the workpiece point cloud data of step 1 and the coarse point cloud completed in the first stage of step 2-1 to obtain Then Perform minimum density sampling as follows:

[0061] The specific implementation process of the minimum density sampling is as follows:

[0062] Will The i-th sampling point is marked as c i , the set of the first i sampling points is represented by C i ={c j |1≤j≤i}. In each iteration, a point c with the minimum density is returned i , the minimum density point can be expressed as:

[0063]

[0064] Where x indicates that it is not in C i-1 For other points within , σ corresponds to the size of the neighborhood considered.

[0065] The residual network is used to refine and complete the output of the minimum density sampling module. The design of the residual network is based on PointNet, which processes a point cloud and outputs a three-channel residual, and then outputs the final point cloud by adding the residual point by point.

[0066] 2-3 Input the point cloud completed in the second stage into the differentiable point cloud discriminator module for adversarial training with the generative network, thereby improving the quality of point cloud generation and better restoring the original details of the workpiece point cloud;

[0067] Because the SGAN network mainly relies on the generator (decoder based on style folding) and the discriminator for adversarial training, so that the point cloud generation effect can be better and better, the discriminator module consists of two parts. The first is a differentiable point cloud renderer that can convert 3D points into 2D depth maps; the second is to use a sequence of 2D convolutional layers with spectral normalization and LeakyReLU activation to identify the output probability, and then conduct adversarial training on the generator and the discriminator.

[0068] Specifically, the differentiable point cloud discriminator module is a discriminator that combines a differentiable renderer and a 2D convolutional layer. Specifically, the point cloud completed in the second stage is first rendered from a 3D point cloud into a 2D depth map through the discriminator, the probability is obtained through the 2D convolutional layer, and then the discriminator and the generator are used for adversarial training.

[0069] The differentiable renderer is as follows: the point cloud P completed in the second stage is projected and rendered from the perspective of the camera posture v. For each 3D point p∈P, the point coordinates of p (p x ,p y ,p z ), after projection rendering, the pixel position is obtained and depth In order to ensure differentiability, the depth Rasterize to get F p , the obtained 2D depth map It can be expressed as:

[0070]

[0071] Where Υ(x) represents the kernel function of Gaussian shape, Υ(x) = exp(-x 2 / 2ɡ 2 ), where ɡ is a hyperparameter that controls the influence radius of the point.

[0072] right The F obtained by rasterization p , which can be expressed as:

[0073]

[0074] F p Represents the negative normalized depth of the point in the range [0,1]. Selecting different features for rendering is also supported, enabling rendering of any point feature without restriction.

[0075] Then render the output 2D depth map Input to a sequence of 2D convolutional layers with spectral normalization and LeakyReLU activation to obtain the probability P r1 .

[0076] In addition, the real point cloud P gt Input to the discriminator and get the corresponding probability P r2 ;

[0077] The probability P of the discriminator output r1 and P r2 Feedback is given to the generating network, so that the generator and the discriminator can be trained adversarially.

[0078] During the training of the entire SGAN network, the total loss It can be expressed as:

[0079]

[0080] in, is the adversarial loss from the differentiable point cloud discriminator module, is the expansion loss, is the reconstruction loss, is the fidelity loss, is the depth map matching loss, is the feature matching loss. w R 、w f 、w adv 、w D 、w feature 、w e Represents the weight of the corresponding loss.

[0081] The mean square error from GAN training is expressed as:

[0082]

[0083] Among them, G represents the decoder based on style folding, P represents the point cloud after the second stage completion, and P gt Represents the real point cloud.

[0084] Reconstruction loss Match the point cloud P completed in the second stage with the real point cloud. It can be expressed as:

[0085]

[0086] Where P c represents the rough point cloud completed in the first stage, P represents the point cloud completed in the second stage, and P gt represents the real point cloud, dEMD Represents the distance difference between two point clouds, d EMD It is expressed as follows:

[0087]

[0088] P and Q refer to two point clouds, i represents the features selected from n points in point cloud P, j represents the features selected from each point in point cloud Q, d(i,j) represents the distance between the i-th feature and the j-th feature, and f ij It represents the cost required to transfer one unit of features from the i-th feature in distribution P to the j-th feature in distribution Q.

[0089] Fidelity loss The structure used to preserve the original residual cloud P0 in the output P, It can be expressed as:

[0090]

[0091] where ||·||2 represents the L2 norm.

[0092] Depth map matching loss It is expressed as:

[0093]

[0094] Where ||·||1 represents the L1 norm; D represents the image depth of the 2D depth map generated by the differentiable renderer module, H represents the image height of the 2D depth map, and W represents the image width of the 2D depth map.

[0095] Feature matching loss Differentiable point cloud discriminator module features The distance exists in the form of:

[0096]

[0097] Among them, H i is the height of the intermediate result image obtained by each 2D convolution layer, W i is the width of the intermediate result image obtained by each 2D convolution layer, F i is the feature of the intermediate result image obtained by each 2D convolution layer, is the re-weighting factor for each feature. i represents the i-th convolutional layer and 4 represents the number of 2D convolutional layers.

[0098] The loss function can be used to calculate the loss of the completed point cloud and the real point cloud. When the loss converges, it means that the generated point cloud is becoming more and more realistic.

[0099] Step 3, analysis of experimental results, in order to verify the effectiveness of the SGAN network, the present invention uses different methods in a self-built workpiece data set for comparison, and evaluates the effectiveness of the SGAN network on the workpiece data set based on three evaluation indicators: chamfer distance (CD), earth mover distance (EMD) and Fréchet distance (FPD). Among them, CD has high calculation efficiency and has been widely used in point cloud evaluation; EMD can better distinguish local details and density distribution; FPD calculates the Fréchet distance between two point cloud sets. This distance takes into account the shape and distribution of the point cloud and provides a global evaluation method. For these three distance indicators, the lower the degree of completion, the better.

[0100] First, let’s compare the chamfer distance (CD) indicator:

[0101]

[0102] From the CD index analysis, except for gears, the method of the present invention has achieved the best performance on other types of workpieces. On gears, it is only 1.1% higher than the optimal ProxyFormer method. On average, the method of the present invention has also achieved the optimal value, which is 6.4% lower than the suboptimal ProxyFormer method.

[0103] Then there is the comparison of the Earth Moving Distance (EMD) indicator:

[0104]

[0105] From the EMD index analysis, except for gears and bolts, the method of the present invention has achieved the best performance on other types of workpieces, and the performance of gears is second only to the best ProxyFormer and the second best GRNet method. In addition, the method of the present invention has achieved the best performance in the final mean, and the value is 0.9% lower than that of the second best ProxyFormer method.

[0106]

[0107] From the analysis of FPD indicators, the method of the present invention has achieved the best or suboptimal performance on all categories of workpieces, and the final average value has reached the best, which is 9.7% lower than the suboptimal PCN method.

[0108] Based on the above quantitative data analysis, it can be seen that the method proposed in the present invention performs excellently in point cloud completion effect on the self-built workpiece dataset.

[0109] Through completion, a good foundation is laid for the subsequent segmentation, pattern recognition and measurement of the workpiece point cloud.

Claims

1. A workpiece point cloud data completion method based on SGAN network, characterized in that: The following steps are involved: Step 1: Obtain workpiece residual cloud data through a structured light camera; Step 2: The workpiece defect cloud data is completed through the SGAN network; The SGAN network includes an SGANNet network, a point cloud refinement module, and a differentiable point cloud discriminator module; The SGANNet network performs the first stage completion on the workpiece residual point cloud data to obtain the first stage completed coarse point cloud; The point cloud refinement module performs second-stage completion on the coarse point cloud data completed in the first stage to obtain a second-stage completed point cloud P; The differentiable point cloud discriminator module performs adversarial training on the point cloud completed in the second stage and the generative network; The differentiable point cloud discriminator module is a discriminator that combines a differentiable renderer with a 2D convolutional layer. Specifically, the point cloud completed in the second stage is first rendered from a 3D point cloud into a 2D depth map by the discriminator, and the probability is obtained by the 2D convolutional layer. Then, the discriminator and the generator are used for adversarial training. The specific implementation process is: The differentiable renderer performs projection rendering on the point cloud P completed in the second stage from the perspective of the camera posture v. For each 3D point p∈P, the point coordinates of p (p x ,p y ,p z ), after projection rendering, the pixel position is obtained and depth In order to ensure differentiability, the depth Rasterize to get F p , the obtained 2D depth map It is expressed as: Where Υ(x) represents the kernel function of Gaussian shape, Υ(x) = exp(-x 2 / 2ɡ 2 ), where ɡ is a hyperparameter that controls the influence radius of the point; F p The negative normalized depth of a point in the range [0,1] is expressed as: Then render the output 2D depth map Input to a sequence of 2D convolutional layers with spectral normalization and LeakyReLU activation to obtain the probability P r1 ; In addition, the real point cloud P gt Input into the differentiable point cloud discriminator and get the corresponding probability P r2 ; The probability P output by the differentiable point cloud discriminator r1 and P r2 Feedback is given to the generative network, so that the generator and the differentiable point cloud discriminator can be trained adversarially.

2. The method according to claim 1, characterized in that: The SGANNet network includes an edge convolution encoder based on a channel space fusion attention mechanism and a decoder based on style folding; The edge convolution encoder based on the channel space fusion attention mechanism includes three pathways and a spatial attention module: The first path specifically uses N workpiece residual defect cloud data P0 as encoder input, and obtains N k directed edges through a k nearest neighbor algorithm; N k directed edges pass through the first multi-layer perceptron F1 to obtain k edge features Edge Features After an average pooling, the output is Then Then pass through the second multi-layer perceptron MLP and activation function sigmoid to get the output The second path is to With edge features Merge and calculate the vector η; then activate the vector η through maximum pooling and ReLu to transform the k edge features Reduced to a new point feature P′0; The third path is to pass the encoder input through the linear layer F3 to get N point output Then The output obtained by the first path The new feature point P'0 obtained by the second path is added to obtain N outputs The spatial attention module takes N outputs The channel-based global maximum pooling and global average pooling are used, and then the two pooling results are spliced ​​together on the channel. The spliced ​​results are processed by the convolution layer to reduce the channel dimension to 1, and then the sigmoid activation function is used to calculate the final feature map P out , denoted as shape code S g ; The decoder based on style folding encodes the shape of the encoder output S g The grid coordinates are input into the generative network at the same time; each layer of the generative network will go through 3 identical style-based folding layers, and then undergo an affine transformation through a linear layer, and then merge to obtain the final output P c , which is the coarse point cloud completed in the first stage.

3. The method according to claim 2, characterized in that: The vector η of the second path in the edge convolution encoder based on the channel space fusion attention mechanism is expressed as: Among them, δ is the sigmoid function, which is used to transform each edge feature Multiply by η to recalibrate it; ° represents the mapping multiplication, F2 represents the second multi-layer perceptron MLP and sigmoid operation; i∈[1,N], j∈[1,k].

4. The method according to claim 2, characterized in that: The style-based folding layers are as follows: First, the grid coordinates P in From the N-dimensional coordinate point S in It is composed of a small batch of point activations through the linear layer to obtain an N-dimensional output S, denoted as h in ;h in Through batch normalization, we get In the same layer of the generative network, the first style-based folding layer coordinate point S in Use the unit square [0,1] 2 , the second to third style-based folding layer coordinate points S in Take the output of the previous style-based folding layer; Shape Code S g After the linear layer, we get two modulation parameters α g and β g ; Combined with the modulation parameter α g and β g Denormalize to get h out ,h out It is expressed as: Finally, h out Updated to the encoder input, it is input again to the encoder based on the channel space fusion attention mechanism to obtain the output P' out .

5. The method according to claim 1, characterized in that: The point cloud refinement module includes a minimum density sampling module and a residual network; The minimum density sampling module combines the workpiece residual point cloud P0 and the rough point cloud P completed in the first stage c Merge Then Perform minimum density sampling; The residual network is used to refine and complete the output of the minimum density sampling module.

6. The method according to claim 1, characterized in that: The total loss of the SGAN network during training It is expressed as: in, is the adversarial loss from the differentiable point cloud discriminator module, is the expansion loss, is the reconstruction loss, is the fidelity loss, is the depth map matching loss, is the feature matching loss, w R 、w f 、w adv 、w D 、w feature 、w e Represents the weight of the corresponding loss.

7. A workpiece point cloud data completion system implementing the method according to any one of claims 1 to 6, characterized in that: include: The data acquisition module is responsible for acquiring the workpiece residual defect cloud data; The data completion module is responsible for completing the workpiece defect cloud data through the SGAN network.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed in a computer, the computer is caused to execute the method according to any one of claims 1 to 6.

9. A computing device comprising a memory and a processor, characterized in that: The memory stores executable codes, and when the processor executes the executable codes, the method according to any one of claims 1 to 6 is implemented.

Images