A large-scale point cloud local constraint attention feature processing method based on a transformer for a 1024*1024 dimension, an electronic device and a storage medium

Abstract

The present application belongs to the technical field of three-dimensional space data processing and deep learning computer vision, and particularly relates to a large-scale point cloud local constraint attention feature processing method based on transformer facing 1024*1024 dimensions, an electronic device and a storage medium. The present application takes 1024*1024 specification three-dimensional dense point cloud as input, and first carries out denoising, normalization and uniform resampling pretreatment on the point cloud; then calculates the relative coordinate offset of the point cloud, and generates learnable position bias coding by using a multi-layer perception machine. The present application constructs a local space neighborhood, embeds the position coding into the key feature to solve the local attention weight, and combines the multi-head mechanism to strengthen the feature representation; stacks the attention modules and constructs a multi-scale feature pyramid by cooperating with the farthest point sampling, and realizes classification or segmentation output according to the task branch. The present application adopts local constraint attention to reduce the calculation complexity, relies on the learnable position coding to capture geometric details, and takes into account the real-time performance and accuracy of the algorithm, and can be adapted to large-scale point cloud perception scenes such as industrial detection and automatic driving.

Description

Technical Field

[0001] This invention belongs to the fields of three-dimensional spatial data processing, deep learning, and computer vision technology. Specifically, it relates to a method for processing local constrained attention features of large-scale point clouds with 1024×1024 dimensions based on transformer, as well as electronic devices and storage media. Background Technology

[0002] With the widespread adoption of 3D perception devices, dense, large-size point cloud data has been widely used in scenarios such as autonomous driving, real-scene reconstruction, and industrial inspection. After conventional acquisition and standardization processing, dense point cloud datasets with a fixed 1024×1024 size are often formed. Existing mainstream point cloud processing solutions are mainly divided into two categories: 1) Lightweight models based on point-by-point convolution and fixed neighborhood aggregation, which rely on manually set geometric priors and cannot adaptively capture the long-range correlations and local fine structures of irregular point clouds. In dense point clouds, detailed features are severely lost; 2) Traditional global attention models, which directly calculate global self-attention for all point sequences. The computational complexity is quadratic. For 1024×1024 ultra-high density point clouds, there are defects such as redundant parameters, extremely high memory consumption, and large inference latency, which cannot meet the requirements of real-time deployment in engineering.

[0003] Existing technologies also have the following technical shortcomings: conventional position encoding often uses fixed sine and cosine encoding or simple coordinate splicing, without considering the local relative spatial relationships of point clouds, resulting in conflicts between coordinate dimensions and feature dimensions, and poor coupling of feature fusion; nearest neighbor sampling often uses fixed radius or fixed number of samples in a crude manner, without combining attention weights to complete dynamic feature selection, resulting in insufficient accuracy in modeling local geometric boundaries, curvature, and spatial topology; in the encoding and decoding downsampling and upsampling process of large-size 1024×1024 point clouds, multi-scale feature transmission is broken, the skip connection fusion mechanism is simple and crude, and deep semantics and shallow details cannot be effectively compatible, ultimately limiting the accuracy of recognition and segmentation.

[0004] In summary, existing technologies struggle to simultaneously balance computational efficiency, feature representation capabilities, and multi-scale information integrity in 1024×1024 dense point cloud scenarios, making it difficult to achieve a balance between engineering feasibility and algorithm accuracy. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies and provide a method, electronic device, and storage medium for local constrained attention feature processing of large-scale point clouds with 1024×1024 dimensions based on transformers. It addresses the technical challenges of existing global attention mechanisms in processing large-scale dense point clouds (1024×1024), including high computational complexity and deployment difficulties; weak spatial structure capture capability of fixed neighborhood modeling; poor adaptability of static position encoding; and fragmented multi-scale feature fusion, leading to low semantic parsing accuracy and poor real-time performance in dense point clouds.

[0006] To achieve the above objectives, this application adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for local constrained attention feature processing of large-scale point clouds with 1024×1024 dimensions based on transformer, characterized by comprising the following steps:

[0008] S01. The original three-dimensional dense point cloud is subjected to outlier denoising, coordinate normalization and uniform resampling in sequence to generate a standard input point cloud with a fixed size of 1024×1024. Each point in the standard input point cloud contains three-dimensional coordinates and an initial shallow feature vector.

[0009] S02. Using the center points within the standard input point cloud as a reference, calculate the three-dimensional relative coordinate offset between the center point and its nearest neighbor points, and generate a dimension-matched learnable position offset encoding vector through multilayer perceptron mapping. This position offset encoding vector is matched with the initial shallow feature vector.

[0010] S03. Construct a local spatial neighborhood by limiting a fixed number of nearest neighbors for each center point, generate query features, key features and value features within the local neighborhood, superimpose the position bias encoding vector onto the key features, and obtain the local attention weights by scaling dot product and Softmax normalization.

[0011] S04. Using the local attention weights to aggregate neighborhood value features, local fusion features are obtained. After layer normalization, feedforward network transformation and residual connection, single-module attention enhancement features are output to complete the calculation of a local constrained attention module.

[0012] S05. Stack several locally constrained attention modules, use farthest point sampling to downsample step by step to extract multi-scale deep coding features, and distinguish the processing branches according to the downstream task type:

[0013] The segmentation task utilizes skip connections to fuse shallow detail features, and then recovers the original 1024×1024 feature map through trilinear interpolation of 3D features, completing point-by-point feature mapping and outputting dense point cloud semantic segmentation results.

[0014] The target classification task enters the feature mapping branch. The classification task abandons the decoding process and directly performs global feature aggregation and mapping on the deep encoded features to output the classification and recognition results.

[0015] Further optimization involves step S01, where coordinate normalization employs the standard mean-standard deviation standardization method, resulting in a normalized coordinate interval of... The initial shallow feature vector dimension satisfies .

[0016] Further optimization: In step S02, the calculation method for the three-dimensional relative coordinate offset is as follows: Let the coordinates of the query center point be... The coordinates of the nearest neighbor points within the neighborhood are The offset is ;

[0017] The multilayer perceptron employs a 2-3 layer fully connected network with ReLU as its activation function.

[0018] The dimension of the learnable positional bias encoding vector is equal to the dimension of the initial shallow feature vector;

[0019] The number of samples from the nearest neighbors satisfies .

[0020] Further optimization involves step S03, where the calculation of the local attention weights employs a multi-head attention mechanism, with the number of multi-head attention satisfying the following... The formula for calculating the single-head local attention weight is as follows:

[0021]

[0022] In the formula, For single-head key feature dimensions, Where Q is the scaling factor and Q is the query feature. is the transpose of the key features.

[0023] Further optimization, in step S05, the stacking number of the local constraint attention modules satisfies Using the 1024×1024 standard input point cloud generated in step S01 as the initial scale, the sizes after step-by-step downsampling through farthest point sampling are 512×512, 256×256, and 128×128 respectively.

[0024] Further optimization: In step S05, the fusion method of the skip connection is a channel splicing method, the expression of which is:

[0025]

[0026] In the formula, For the deep semantic features of the decoder, These are the shallow detail features of the corresponding layer of the encoder.

[0027] Further optimization involves the number of upsampling stages in the 3D feature trilinear interpolation corresponding to the number of downsampling stages, ultimately resulting in a forced constraint on the output feature scale. .

[0028] Further optimization is achieved in step S05, where the decoding process is omitted when the downstream task is point cloud target classification; global max pooling is performed on the multi-scale deep encoded features to obtain the global feature vector, which is then mapped through a fully connected layer to output the classification result.

[0029] Further optimization is achieved in step S05, where the downstream task is point cloud semantic segmentation. After the feature map is restored, a 1×1 convolution is used to perform point-by-point feature mapping on the 1024×1024 complete feature map, and the semantic classification result of each point is output.

[0030] In a second aspect, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of a condition-guided intelligent control method for assembly quality.

[0031] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a computer processor, implements any step of a condition-guided intelligent control method for assembly quality.

[0032] The beneficial effects achieved by this technical solution are:

[0033] 1. Significantly reduce the overhead of dense point cloud computing: Abandoning global full attention calculation, the algorithm adopts the k-nearest neighbor local constraint mechanism, which reduces the cloud computing complexity of 1024×1024 large-size points from O(N²) to O(N·k), significantly reducing memory usage and computation. The algorithm can be directly deployed on mid-range computing devices, balancing high precision and real-time performance.

[0034] 2. Enhance 3D spatial geometric modeling capabilities: Introduce learnable relative position offset coding to replace traditional fixed coding, adaptively fitting the spatial distribution patterns of dense point clouds; Combine local neighborhood attention weighting to accurately capture local curvature, edges, and topological details of point clouds, solving the problem of loss of structural features in large-size dense point clouds.

[0035] 3. Complete coupling of multi-scale features: The architecture of step-down sampling encoding + skip connection upsampling decoding is adopted to ensure the uniformity of 1024×1024 input and output dimensions. Shallow detail features and deep semantic features are fused in both directions, which greatly improves the overall accuracy and robustness of dense point cloud segmentation and classification tasks.

[0036] 4. Strong versatility and engineering adaptability: The overall methodology is standardized and the parameters can be dynamically configured. It is uniformly adapted to dense point clouds of fixed specifications of 1024×1024 and can be seamlessly migrated to various downstream engineering scenarios such as autonomous driving perception, industrial 3D inspection, and scene reconstruction. Attached Figure Description

[0037] Figure 1 This is a diagram of the overall algorithm architecture.

[0038] Figure 2 This is a schematic diagram of the internal structure of the local constraint attention module;

[0039] Figure 3 Diagram of the multi-scale encoding / decoding and skip connection fusion architecture;

[0040] Figure 4 This image shows a comparison of the image segmentation results of similar algorithms used in downstream tasks. Detailed Implementation

[0041] The implementation of the present invention will be further described below with reference to the accompanying drawings.

[0042] A transformer-based method for local constraint attention feature processing of large-scale point clouds with 1024×1024 dimensions includes the following steps:

[0043] S01. Point Cloud Preprocessing and Standard Input Construction: The acquired original 3D dense point cloud is sequentially subjected to outlier denoising, coordinate normalization, and uniform resampling to generate a standard input point cloud with a fixed size of 1024×1024. Each point in the standard input point cloud contains 3D coordinates and an initial shallow feature vector. Coordinate normalization uses a standard mean-standard deviation normalization method, resulting in a coordinate range of [-1, 1]. The initial shallow feature vector dimension meets the preset dimension requirement, i.e. ;

[0044] S02. Learnable Position Offset Encoding Generation: Using the center points within the standard input point cloud as a reference, calculate the three-dimensional relative coordinate offset between the center point and its nearest neighbors. Let the coordinates of the query center point be... The coordinates of the nearest neighbor points within the neighborhood are The offset is ;

[0045] A dimension-matched learnable location bias encoding vector is generated by mapping using a multilayer perceptron. This multilayer perceptron employs a 2-3 layer fully connected network with ReLU activation function. The dimension of the learnable location bias encoding vector is equal to the dimension of the initial shallow feature vector, and the number of nearest neighbor samples meets a preset range. ;

[0046] S03. Local Attention Weight Calculation: A fixed number of nearest neighbors are defined for each center point to construct a local spatial neighborhood. Query features, key features, and value features are generated within this local neighborhood. The positional bias encoding vector is then superimposed onto the key features. This step employs a multi-head attention mechanism to calculate the local attention weights. The number of heads meets a preset value range. The formula for calculating the single-head local attention weight is as follows:

[0047]

[0048] In the formula, For single-head key feature dimensions, Where Q is the scaling factor and Q is the query feature. is the transpose of the key features;

[0049] S04. Single-module attention feature enhancement: The local attention weights are used to aggregate neighborhood value features to obtain local fusion features. After layer normalization, feedforward network transformation and residual connection, single-module attention enhancement features are output.

[0050] S05, Multi-scale Coding and Task Branching: A preset number of local constraint attention modules are stacked in series, and the number of stacked local constraint attention modules meets a preset value range. The modules are downsampled stepwise through farthest point sampling, using a standard input point cloud of 1024×1024 as the initial scale. The sizes after stepwise downsampling are 512×512, 256×256, and 128×128, respectively, compressing the spatial dimension and generating multi-scale deep coding features. Different processing branches are divided according to the downstream task type. For point cloud semantic segmentation tasks, the decoding upsampling branch is used, and for point cloud target classification tasks, the global feature mapping branch is used to adapt to different downstream visual recognition tasks.

[0051] Semantic segmentation task decoding processing: When the downstream task is point cloud semantic segmentation, based on the skip connection fusion encoding stage multi-scale shallow detail features, the skip connection fusion method is channel concatenation, and its expression is:

[0052]

[0053] In the formula, For the deep semantic features of the decoder, The shallow detail features of the encoder's corresponding layer are used; the three-dimensional features are upsampled step by step through trilinear interpolation, with the upsampling level corresponding to the downsampling level, and the spatial geometric information is completed layer by layer to restore a complete feature map of size 1024×1024; subsequently, 1×1 convolution is used to perform point-by-point feature mapping on the complete feature map, and the semantic classification result of each point is output.

[0054] Feature mapping for target classification task: When the downstream task is point cloud target classification task, the multi-scale deep coding features generated in step S05 are subjected to global max pooling to obtain a global feature vector, which is then mapped through a fully connected layer to output the point cloud category recognition result.

[0055] Example 1

[0056] Processing methods for target classification in 1024×1024 dense point clouds

[0057] This embodiment is used for large-scale 3D workpiece point cloud classification. The overall system architecture diagram is as follows: Figure 1 As shown, the specific steps are as follows:

[0058] 1) Collect the original three-dimensional dense point cloud of the industrial workpiece, remove noise from distant outliers, normalize the coordinates to the [-1,1] interval, and uniformly resample to generate a 1024×1024 standard input point cloud. The initial shallow feature dimension of each point is set to 64 dimensions.

[0059] 2) Based on the center points of each point in the standard input point cloud, the number of nearest neighbors of each point is k=16. A 2-layer MLP is used to map the relative coordinate offset to generate a 64-dimensional learnable position offset encoding vector.

[0060] 3) Construct a local spatial neighborhood based on a preset fixed number of nearest neighbors, and superimpose the learnable positional bias encoding vector generated in step 2 onto the key feature K to complete the positional information injection; perform a linear mapping of query feature Q, key feature K, and value feature V within the neighborhood, calculate the scaled dot product attention and obtain the local attention weights through Softmax normalization; use these weights to aggregate the value feature V, and then perform layer normalization, feedforward network transformation, and residual connection to output the enhanced features of the current local constraint attention module; in this embodiment, the number of multi-head attention heads is set to 4, and the detailed structure of the local constraint attention module is as follows: Figure 2 As shown;

[0061] We utilize local attention weights to aggregate neighborhood value features, and combine this with layer normalization, feedforward network, and residual connection to obtain single-module attention enhancement features.

[0062] 4) Six groups of local constraint attention modules are stacked in series. Farthest point sampling is performed once for every two groups of modules, downsampling sequentially to 512×512 and 256×256. A detailed diagram of the multi-scale coding architecture is shown below. Figure 3 As shown;

[0063] 5) This embodiment is for a target classification task. The decoding upsampling process is abandoned. Global max pooling is performed on the final deep features at the end of the encoding to obtain a global feature vector. Then, a fully connected classification layer is concatenated to complete the feature mapping and output the industrial workpiece category recognition result.

[0064] 6) Layer normalization and residual connections are added throughout the process to suppress gradient degradation in deep networks, ensure the stability of large point cloud inference, and improve the robustness of the model.

[0065] Example 2

[0066] A processing method for semantic segmentation of 1024×1024 indoor dense point clouds

[0067] This embodiment is used for 3D semantic segmentation of large-scene indoor environments. The overall system architecture diagram is as follows: Figure 1 As shown, the specific steps are as follows:

[0068] 1) The indoor panoramic dense point cloud is filtered, noise-reduced, and density-equalized, and uniformly normalized to 1024×1024 dimensions, with the initial feature dimension of a single point set to 128 dimensions;

[0069] 2) With a single center point having a nearest neighbor number of k=24, a 3-layer deep MLP is used to complete the relative coordinate mapping and output a 128-dimensional adaptive position offset code.

[0070] 3) This module is the core module for implementing the large-scale point cloud local constraint attention feature processing method. A detailed module design structure diagram is shown below. Figure 2 As shown, in the point cloud semantic segmentation settings, the number of multi-head attention heads is set to 8 to enhance the discriminative power of fine-grained features, and neighborhood attention weighted aggregation of local structural information.

[0071] We utilize local attention weights to aggregate neighborhood value features, and combine this with layer normalization, feedforward network, and residual connection to obtain single-module attention enhancement features.

[0072] 4) Stack 8 sets of local constraint attention modules, downsampling stepwise to 512×512, 256×256, and 128×128 to construct a multi-scale semantic feature pyramid. A detailed diagram of the multi-scale coding architecture is shown below. Figure 3 As shown;

[0073] 5) The decoder performs trilinear interpolation upsampling step by step, and fuses the detailed features of each level of the encoder through skip connections, finally restoring a full-size feature map of 1024×1024;

[0074] 6) Point-by-point semantic classification is achieved using 1×1 convolution, outputting a multi-class semantic segmentation map. To demonstrate the effectiveness of this algorithm for downstream tasks, a comparison with similar algorithms (the comparison algorithm is PointNet) was performed. Specific visualization comparison results are shown below. Figure 4 As shown; at the same time, conditional random fields are introduced for post-processing to smooth object boundaries and eliminate noise and spiking in dense point cloud segmentation.

[0075] Example 3

[0076] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of a transformer-based method for local constraint attention feature processing of large-scale point clouds with 1024×1024 dimensions as described in Embodiment 1.

[0077] Furthermore, the method for local constraint attention feature processing of large-scale point clouds with 1024×1024 dimensions based on transformers described in Embodiment 1 can be implemented as a computer software program. For example, this embodiment includes a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the method. In such an embodiment, the computer program can be downloaded and installed from a network, and / or installed from a removable medium. When the computer program is executed by a processor, it performs the functions defined in the method of this application.

[0078] Example 4

[0079] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements any step of a method for implementing a large-scale point cloud local constraint attention feature processing method based on transformer and oriented to 1024×1024 dimensions, as described in Embodiment 1.

[0080] The computer-readable medium described in this application may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof. The computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages—such as Python and C++—as well as conventional procedural programming languages ​​or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0081] In this embodiment, the computer-readable storage medium can be accelerated using hardware such as a GPU. The parallel computing advantage of the GPU is used to accelerate any step in the method for local constraint attention feature processing of large-scale point clouds with 1024×1024 dimensions based on transformer as described in Embodiment 1.

[0082] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

[0083] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for local constraint attention feature processing of large-scale point clouds with 1024×1024 dimensions based on transformer, characterized in that, Includes the following steps: S01. The original three-dimensional dense point cloud is subjected to outlier denoising, coordinate normalization and uniform resampling in sequence to generate a standard input point cloud with a fixed size of 1024×1024. Each point in the standard input point cloud contains three-dimensional coordinates and an initial shallow feature vector. S02. Using the center points within the standard input point cloud as a reference, calculate the three-dimensional relative coordinate offset between the center point and its nearest neighbor points, and generate a dimension-matched learnable position offset encoding vector through multilayer perceptron mapping. This position offset encoding vector is matched with the initial shallow feature vector. S03. Construct a local spatial neighborhood by limiting a fixed number of nearest neighbors for each center point, generate query features, key features and value features within the local neighborhood, superimpose the position bias encoding vector onto the key features, and obtain the local attention weights by scaling dot product and Softmax normalization. S04. Using the local attention weights to aggregate neighborhood value features, local fusion features are obtained. After layer normalization, feedforward network transformation and residual connection, single-module attention enhancement features are output to complete the calculation of a local constrained attention module. S05. Stack several locally constrained attention modules, use farthest point sampling to downsample step by step to extract multi-scale deep coding features, and distinguish the processing branches according to the downstream task type: The semantic segmentation task enters the decoding branch. The segmentation task uses skip connections to fuse shallow detail features. After three-dimensional feature trilinear interpolation, it recovers the original 1024×1024 feature map by step-by-step upsampling, completes point-by-point feature mapping, and outputs dense point cloud semantic segmentation results. The target classification task enters the feature mapping branch. The classification task abandons the decoding process and directly performs global feature aggregation and mapping on the deep encoded features to output the classification and recognition results.

2. The processing method according to claim 1, characterized in that, In step S01, coordinate normalization uses the standard mean-standard deviation standardization method, and the normalized coordinate interval is: The initial shallow feature vector dimension satisfies .

3. The processing method according to claim 1, characterized in that, In step S02, the calculation method for the three-dimensional relative coordinate offset is as follows: Let the coordinates of the query center point be... The coordinates of the nearest neighbor points within the neighborhood are The offset is ; The multilayer perceptron employs a 2-3 layer fully connected network with ReLU as its activation function. The dimension of the learnable positional bias encoding vector is equal to the dimension of the initial shallow feature vector; The number of samples from the nearest neighbors satisfies .

4. The processing method according to claim 1, characterized in that, In step S03, the calculation of the local attention weights adopts a multi-head attention mechanism, and the number of multi-head attention satisfies the following conditions: The formula for calculating the single-head local attention weight is as follows: In the formula, For single-head key feature dimensions, Where Q is the scaling factor and Q is the query feature. is the transpose of the key features.

5. The processing method according to claim 1, characterized in that, In step S05, the number of stacked local constraint attention modules satisfies Using the 1024×1024 standard input point cloud generated in step S01 as the initial scale, when the downstream task is classification, six local constraint attention modules are stacked, and downsampled stepwise to 512×512 and 256×256 through farthest point sampling; when the downstream task is segmentation, eight modules are stacked, and downsampled stepwise to 512×512, 256×256, and 128×128.

6. The processing method according to claim 1, characterized in that, In step S05, when the downstream task is point cloud target classification, the decoding process is omitted; global max pooling is performed on the multi-scale deep encoded features to obtain the global feature vector, and then the classification result is output through a fully connected layer.

7. The processing method according to claim 1, characterized in that, In step S05, when the downstream task is point cloud semantic segmentation, after the feature map is restored, a 1×1 convolution is used to perform point-by-point feature mapping on the 1024×1024 complete feature map, and the semantic classification result of each point is output.

8. The processing method according to claim 1, characterized in that, In step S05, the fusion method of the skip connection is channel splicing, and its expression is: In the formula, For the deep semantic features of the decoder, For shallow detail features of the corresponding level of the encoder; and / or, The number of upsampling levels in the three-dimensional feature trilinear interpolation corresponds to the number of downsampling levels, and the final output feature scale is forcibly constrained. .

9. An electronic device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements any step in the condition-guided intelligent control method for assembly quality as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a computer processor, implements any step in the condition-guided intelligent control method for assembly quality as described in any one of claims 1-8.

Images