Weakly supervised point cloud semantic segmentation method and system based on multi-scale feature extraction and classification

Abstract

The present application discloses a weakly supervised point cloud semantic segmentation method and system for multi-scale feature extraction and classification, which belongs to the field of deep learning and includes: obtaining target point cloud data; performing feature extraction and aggregation processing on the target point cloud data according to a preset ERFA module to generate feature maps with different dimensions; querying and interpolating each feature map with different dimensions respectively, and performing feature splicing processing after obtaining the semantic features corresponding to each feature map to generate target semantic features; training the preset initial prediction network model according to the preset multi-scale classification strategy and local contrast regularization strategy to obtain the target prediction network model and generate classification results. The present invention takes the ERFA module as the core, realizes the expansion and enhancement of local area features, and combines multi-scale classification and local contrast regularization strategies to make full use of the semantic information in sparsely labeled data, thereby improving the accuracy and precision of classification.

Description

Technical Field

[0001] This application belongs to the field of deep learning technology, and specifically relates to a weakly supervised point cloud semantic segmentation method and system for multi-scale feature extraction and classification. Background Art

[0002] Semantic segmentation of large-scale 3D point clouds is a fundamental task. By analyzing the semantics of spatial point clouds, it helps AI machines understand our complex 3D world. It plays a vital role in diverse fields, including autonomous driving, robot navigation, smart buildings, and cultural heritage preservation. Compared to previous methods based on manual feature extraction, deep learning-based approaches have demonstrated significant advantages, automatically learning complex feature representations to improve segmentation accuracy. However, these deep learning methods often rely heavily on large amounts of fully labeled data for model training. When faced with datasets containing millions or even tens of millions of point clouds, fully labeled methods become impractical. On the one hand, labeling large point clouds is expensive and time-consuming; on the other hand, the labeling process is prone to errors, which can affect the model's generalization ability. Inspired by the successful application of weakly supervised methods in 2D image processing, researchers have begun to focus on training 3D point cloud semantic segmentation models using fewer point cloud labels.

[0003] In recent years, weakly supervised segmentation methods have been continuously developed. Existing methods are mainly based on limited 3D label-based methods. These methods have shown significant advantages on some small-scale datasets, but still have some limitations, including the following aspects:

[0004] First, existing methods often focus on single-scale classification of point clouds. Single-scale classification may have certain advantages in small-scale datasets, but in large-scale point cloud datasets, methods combined with weak labels are unable to cope. Large-scale scenes are more complex, contain a large number of spatial point clouds, and the foreground and background scales of the same scene vary significantly. For example, in indoor scenes, foreground objects such as furniture are smaller in scale, while background objects such as walls and ceilings are larger in scale. Single-scale classification is insufficient to capture local and global structural features, resulting in limited performance of the model when processing complex scenes and low classification accuracy.

[0005] Second, existing methods still fail to adequately consider local semantic information. In weakly supervised semantic segmentation, local semantic information is crucial for improving segmentation accuracy. However, existing methods often fail to fully utilize local semantic information, resulting in insufficient perception of local details. For example, in some scenarios, the boundary regions between different categories may contain rich local semantic information, but existing methods fail to effectively utilize this information, thus affecting segmentation accuracy.

[0006] Third, existing methods fail to fully utilize the valuable information in weakly labeled data. Weakly labeled data often contains noise and uncertainty, but existing methods fail to effectively address these issues. For example, some methods directly use weakly labeled data as supervisory signals, ignoring the impact of label noise on model training. This results in inaccurate scene segmentation. Summary of the Invention

[0007] The purpose of the embodiments of the present application is to provide a weakly supervised point cloud semantic segmentation method and system for multi-scale feature extraction and classification, which can solve the problem of low classification accuracy and precision in the existing technology.

[0008] In order to solve the above technical problems, this application is implemented as follows:

[0009] In a first aspect, an embodiment of the present application provides a weakly supervised point cloud semantic segmentation method with multi-scale feature extraction and classification, the method comprising:

[0010] Acquire and preprocess the original point cloud data to generate target point cloud data, which consists of the number, position features and semantic features of the point cloud;

[0011] Perform feature extraction and feature aggregation on the target point cloud data according to the preset ERFA module to generate feature maps with different dimensions;

[0012] Query and interpolate the feature maps of each dimension respectively, obtain the semantic features corresponding to each feature map, and then perform feature splicing to generate the target semantic features. The target semantic features include the semantic features corresponding to the query point cloud and the semantic features corresponding to the neighboring point clouds of the query point cloud.

[0013] The preset initial prediction network model is trained according to the preset multi-scale classification strategy and local contrast regularization strategy to obtain the target prediction network model;

[0014] The target semantic features are input as training signals into the target prediction network model for processing to generate classification results.

[0015] As an optional implementation scheme of the first aspect of the present application, the preset ERFA module specifically includes: a first shared MLP layer, a second shared MLP layer, a third shared MLP layer, an expanded feature extraction layer, and an activation function layer, wherein the second shared MLP layer, the expanded feature extraction layer, the third shared MLP layer, and the activation function layer are connected in series from left to right, and the first shared MLP layer is connected in parallel with the second shared MLP layer, the expanded feature extraction layer, the third shared MLP layer, and the activation function layer.

[0016] As an optional implementation of the first aspect of the present application, the specific process of performing feature extraction and feature aggregation processing on the target point cloud data according to the preset ERFA module to generate multiple feature maps with different dimensions includes:

[0017] Performing feature extraction and feature aggregation processing on the target point cloud data according to the first ERFA module to generate a first feature map;

[0018] Performing feature extraction and feature aggregation processing on the first feature map according to the second ERFA module to generate a second feature map;

[0019] performing feature extraction and feature aggregation processing on the second feature map according to the third ERFA module to generate a third feature map;

[0020] performing feature extraction and feature aggregation processing on the third feature map according to the fourth ERFA module to generate a fourth feature map;

[0021] Among them, the semantic feature dimensions of the first feature map, the second feature map, the third feature map, and the fourth feature map are the same, the position feature dimension of the fourth feature map is higher than that of the third feature map, the position feature dimension of the third feature map is higher than that of the second feature map, the position feature dimension of the second feature map is higher than that of the first feature map, the number of point clouds of the fourth feature map is lower than that of the third feature map, the number of point clouds of the third feature map is lower than that of the second feature map, and the number of point clouds of the second feature map is lower than that of the first feature map.

[0022] As an optional implementation of the first aspect of the present application, the specific steps of performing feature extraction and feature aggregation processing on the target point cloud data according to the first ERFA module to generate a first feature map include:

[0023] Performing dimension transformation processing on the target point cloud data according to the first shared MLP layer and the second shared MLP layer respectively, and correspondingly obtaining a first semantic feature after dimension transformation and a second semantic feature after dimension transformation;

[0024] Performing position encoding processing on the second semantic feature after dimension transformation to generate input point cloud features;

[0025] Performing expansion feature extraction processing on the input point cloud features to generate a first aggregated feature;

[0026] Performing dimension transformation on the first aggregated feature according to the third shared MLP layer and concatenating the first semantic feature after dimension transformation to obtain a concatenated feature;

[0027] The concatenated features are processed according to an activation function to generate a first feature map.

[0028] As an optional implementation of the first aspect of the present application, the specific process of obtaining the target semantic feature includes:

[0029] Process the input point cloud features to obtain point cloud features with a point cloud quantity of 1, a position feature of 3 dimensions, and a semantic feature of d dimensions;

[0030] Split the point cloud features with a point cloud number of 1, a position feature of 3 dimensions, and a semantic feature of d dimensions to obtain point cloud features with a point cloud number of 1, a position feature of 3 dimensions, and a semantic feature of 0 dimensions, as well as point cloud features with a point cloud number of 1, a position feature of 0 dimensions, and a semantic feature of d dimensions;

[0031] The first KNN processing is performed on the point cloud features with 1 point cloud, 3-dimensional position features, and 0-dimensional semantic features, as well as the point cloud features with 1 point cloud, 0-dimensional position features, and d-dimensional semantic features, respectively, to obtain the point cloud features with k1 point cloud, 3-dimensional position features, and 0-dimensional semantic features, and the point cloud features with k1 point cloud, 0-dimensional position features, and d-dimensional semantic features;

[0032] Perform position encoding processing on the point cloud features with k1 point clouds, 3-dimensional position features, and 0-dimensional semantic features to obtain position encoding features with k1 point clouds and d-dimensional position features;

[0033] Perform the second KNN processing on the point cloud feature with 1 point cloud, 3-dimensional position feature, and d-dimensional semantic feature to obtain the point cloud feature with k2 point clouds, 3-dimensional position feature, and d-dimensional semantic feature. The focus scales of the first KNN processing and the second KNN processing are different.

[0034] The point cloud features with k2 point clouds, 3-dimensional position features and d-dimensional semantic features are split to obtain point cloud features with k2 point clouds, 3-dimensional position features and 0-dimensional semantic features and point cloud features with k2 point clouds, 0-dimensional position features and d-dimensional semantic features;

[0035] Perform position encoding processing on the point cloud features with k2 point clouds, 3-dimensional position features, and 0-dimensional semantic features to obtain position encoding features with k2 point clouds and d-dimensional position features;

[0036] The position coding features with k2 point clouds and d-dimensional position features are concatenated with the position coding features with k1 point clouds and d-dimensional position features, and then subjected to attention pooling to obtain enhanced position coding features.

[0037] The point cloud features with k1 points, 0-dimensional position features, and d-dimensional semantic features are concatenated with the point cloud features with k2 points, 0-dimensional position features, and d-dimensional semantic features, and then subjected to attention pooling processing to obtain enhanced point cloud features;

[0038] The enhanced position encoding features and the enhanced point cloud features are concatenated to obtain the target semantic features.

[0039] As an optional implementation of the first aspect of the present application, the local contrast regularization strategy includes:

[0040]

[0041]

[0042] in, represents the local contrast regularization strategy loss function, represents the number of points in the point cloud, represents the hyperparameters set, represents the number of nearest neighbor points, Represents the original point cloud Through the function The predicted value obtained after processing, Indicates parameters, Represents a query point The predicted value of represents the mean of the predicted values, Represents the indicator function.

[0043] As an optional implementation of the first aspect of the present application, the multi-scale classification strategy includes:

[0044]

[0045] in, represents the multi-scale classification strategy loss function, is the total number of labeled classes, represents the number of points in the point cloud, is the total number of layers, For The true value of the class, For Class The predicted value of the layer.

[0046] In a second aspect, an embodiment of the present application provides a weakly supervised point cloud semantic segmentation system with multi-scale feature extraction and classification, the system comprising:

[0047] The preprocessing module is used to acquire and preprocess the original point cloud data to generate target point cloud data, which consists of the number, position features and semantic features of the point cloud;

[0048] The ERFA module is used to perform feature extraction and feature aggregation on the target point cloud data to generate feature maps with multiple dimensions;

[0049] The target semantic feature generation module is used to query and interpolate the feature maps of different dimensions respectively, obtain the semantic features corresponding to each feature map, and then perform feature splicing processing to generate the target semantic features. The target semantic features include the semantic features corresponding to the query point cloud and the semantic features corresponding to the neighboring point clouds of the query point cloud;

[0050] A target preset network model generation module is used to train the preset initial prediction network model according to the preset multi-scale classification strategy and local contrast regularization strategy to obtain the target prediction network model;

[0051] The classification module is used to input the target semantic features as training signals into the target prediction network model for processing to generate classification results.

[0052] In a third aspect, an embodiment of the present application provides an electronic device comprising a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the method of the first aspect.

[0053] In a fourth aspect, an embodiment of the present application provides a readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the method of the first aspect are implemented.

[0054] Compared with the prior art, the weakly supervised point cloud semantic segmentation method with multi-scale feature extraction and classification provided by the present invention has the following advantages:

[0055] First, an ERFA module is proposed. The ERFA module can maximize the feature representation of local areas while ensuring feature generalization performance. It focuses on the feature aggregation method from two different scale regions. The main operation of the ERFA module is to expand feature extraction and solve the network degradation problem by adding residual connections.

[0056] Second, we utilize the attention mechanism to design an expanded feature extraction layer that considers the aggregation of local features at different scales as well as information in high-dimensional and low-dimensional feature spaces.

[0057] Third, in order to obtain more stable features from disordered point cloud data, this section uses local contrast regularization to perform comparative learning on the local information of original and enhanced data, fully considering the strong local semantic homogeneity of large-scale scenes.

[0058] Fourth, after processing point clouds of different scales, multi-scale features are obtained, and the features of different scales are extracted and analyzed, and finally sent to the classifier to produce the final prediction results, thereby improving the accuracy and precision of the classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flowchart of the weakly supervised point cloud semantic segmentation method with multi-scale feature extraction and classification provided by the first embodiment of the present application;

[0060] Figure 2 This is an overall structural diagram of the multi-scale feature extraction and classification provided by the first embodiment of the present application;

[0061] Figure 3 is a structural diagram of the expanded feature extraction layer provided in the first embodiment of the present application;

[0062] Figure 4 is an internal structure diagram of the ERFA module provided in the first embodiment of the present application;

[0063] Figure 5 This is a diagram of the internal structure of a weakly supervised point cloud semantic segmentation system with multi-scale feature extraction and classification provided in the second embodiment of the present application. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0065] The terms "first," "second," and the like in the specification and claims of this application are used to distinguish similar objects and are not used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate so that the embodiments of this application can be implemented in an order other than those illustrated or described herein. In addition, the term "and / or" in the specification and claims refers to at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0066] The following, in conjunction with the accompanying drawings, describes in detail the weakly supervised point cloud semantic segmentation method and system for multi-scale feature extraction and classification provided by the embodiments of the present application through specific embodiments and their application scenarios.

[0067] Example 1

[0068] See also Figure 1 , which represents a flow chart of the weakly supervised point cloud semantic segmentation method with multi-scale feature extraction and classification provided by the present invention, including steps S1 to S5.

[0069] Step S1: Acquire and preprocess the original point cloud data to generate target point cloud data, which consists of the number, position features and semantic features of the point cloud.

[0070] Specifically, the present invention performs data enhancement processing on the original point cloud data to obtain target point cloud data. The target point cloud data consists of the number, position features and semantic features of the point cloud. The mathematical expression is:

[0071]

[0072] in, represents a point cloud, represents the number of point clouds, Representing point clouds The location characteristics, Representing point clouds The semantic features of .

[0073] See also Figure 2 , which represents the overall structure diagram of the multi-scale feature extraction and classification provided by the present invention. The overall structure diagram can be divided into three parts: encoding, point cloud semantic query and decoding, as shown below:

[0074] (1) Coding;

[0075] The encoder's core function is to integrate features from the raw point cloud to facilitate subsequent segmentation tasks. The encoder's multi-layered architecture captures the diverse receptive fields of the point cloud: the upper layers focus on preserving fine-grained local features, while the lower layers integrate broader, long-range correlated features. Each encoder layer consists of a designed ERFA module and random sampling operations: the ERFA module aggregates local features, while the random sampling operation downsamples the point cloud. These two components work together to effectively extract hierarchical features within the encoder.

[0076] (2) Point cloud semantic query;

[0077] In the weakly supervised semantic segmentation task of point clouds, due to the limited number of labeled point clouds, Figure 2The semantic query module shown in the middle part is designed to efficiently utilize as many relevant features as possible. To query the center, search the nearby point cloud based on Euclidean distance The semantic features of and The semantic feature combination of is used as the training signal to input into the model. The workflow of the semantic query module consists of two key steps: first, searching for neighboring features of the query point cloud, and second, performing feature interpolation between the query point cloud and the neighboring point clouds.

[0078] (3) Decoding;

[0079] The decoder design is based on a U-shaped segmentation structure and processes features for a specific labeled point cloud, eliminating the need for interpolation and restoring the original feature distribution map. During the decoding process, features at different levels are first processed by a semantic query module, then integrated and concatenated at the channel level to ultimately achieve point cloud prediction.

[0080] Step S2: Perform feature extraction and feature aggregation processing on the target point cloud data according to the preset ERFA module to generate multiple feature maps with different dimensions.

[0081] Specifically, see Figure 4 , represents the internal structure of the ERFA (Enhanced Regional Feature Aggregation, ERFA) module provided by the present invention. ERFA represents the extended regional feature aggregation unit. Within the extended regional feature aggregation unit, the present invention comprehensively considers the aggregation of local features of different scales as well as information in high- and low-dimensional feature spaces. Therefore, the ERFA module does not require additional local feature processing and only includes residual connections to address the problem of vanishing gradients. The internal structure of the ERFA module includes multiple multi-layer perceptron (MLP) layers, specifically including: a first shared MLP layer, a second shared MLP layer, a third shared MLP layer, an expanded feature extraction layer, and an activation function layer. The second shared MLP layer, the expanded feature extraction layer, the third shared MLP layer, and the activation function layer are connected serially from left to right, and the first shared MLP layer is connected in parallel with the second shared MLP layer, the expanded feature extraction layer, the third shared MLP layer, and the activation function layer.

[0082] Furthermore, the specific process of performing feature extraction and feature aggregation processing on the target point cloud data according to the preset ERFA module to generate multiple feature maps with different dimensions includes: performing feature extraction and feature aggregation processing on the target point cloud data according to the first ERFA module to generate a first feature map; performing feature extraction and feature aggregation processing on the first feature map according to the second ERFA module to generate a second feature map; performing feature extraction and feature aggregation processing on the second feature map according to the third ERFA module to generate a third feature map; performing feature extraction and feature aggregation processing on the third feature map according to the fourth ERFA module to generate a fourth feature map;

[0083] Among them, the first characteristic map, the second characteristic map, the third characteristic map, and the fourth characteristic map are respectively represented as 、 、 、 ,in Represents the number of point clouds. It can be seen that the semantic feature dimensions of the first, second, third, and fourth feature maps are the same, all 3. The positional feature dimension of the fourth feature map is 512, which is higher than the positional feature dimension 256 of the third feature map. The positional feature dimension 256 of the third feature map is higher than the positional feature dimension 128 of the second feature map. The positional feature dimension 128 of the second feature map is higher than the positional feature dimension 32 of the first feature map. The number of point clouds of the fourth feature map is , which is lower than the number of point clouds in the third feature map , the number of point clouds in the third feature map The number of point clouds below the second feature map , the number of point clouds in the second feature map The number of point clouds below the first feature map .

[0084] Furthermore, the process of performing feature extraction and feature aggregation processing on the target point cloud data according to the first ERFA module to generate a first feature map includes: performing dimensionality transformation processing on the target point cloud data according to the first shared MLP layer and the second shared MLP layer respectively, and correspondingly obtaining the first semantic feature after dimensionality transformation and the second semantic feature after dimensionality transformation; performing position encoding processing on the second semantic feature after dimensionality transformation to generate the input point cloud feature; performing expansion feature extraction processing on the input point cloud feature to generate the first aggregate feature; performing dimensionality transformation processing on the first aggregate feature according to the third shared MLP layer and then performing feature splicing with the first semantic feature after dimensionality transformation to obtain the spliced ​​feature; processing the spliced ​​feature according to the activation function to generate the first feature map.

[0085] The expanded feature extraction layer provided by the present invention is used to obtain target semantic features. The specific process includes the following steps:

[0086] Process the input point cloud features to obtain point cloud features with a point cloud quantity of 1, a position feature of 3 dimensions, and a semantic feature of d dimensions;

[0087] Split the point cloud features with a point cloud number of 1, a position feature of 3 dimensions, and a semantic feature of d dimensions to obtain point cloud features with a point cloud number of 1, a position feature of 3 dimensions, and a semantic feature of 0 dimensions, as well as point cloud features with a point cloud number of 1, a position feature of 0 dimensions, and a semantic feature of d dimensions;

[0088] The first KNN processing is performed on the point cloud features with 1 point cloud, 3-dimensional position features, and 0-dimensional semantic features, as well as the point cloud features with 1 point cloud, 0-dimensional position features, and d-dimensional semantic features, respectively, to obtain the point cloud features with k1 point cloud, 3-dimensional position features, and 0-dimensional semantic features, and the point cloud features with k1 point cloud, 0-dimensional position features, and d-dimensional semantic features, respectively. KNN represents the k-Nearest Neighbor algorithm (KNN).

[0089] Perform position encoding processing on the point cloud features with k1 point clouds, 3-dimensional position features, and 0-dimensional semantic features to obtain position encoding features with k1 point clouds and d-dimensional position features;

[0090] Perform the second KNN processing on the point cloud feature with 1 point cloud, 3-dimensional position feature, and d-dimensional semantic feature to obtain the point cloud feature with k2 point clouds, 3-dimensional position feature, and d-dimensional semantic feature. The focus scales of the first KNN processing and the second KNN processing are different.

[0091] The point cloud features with k2 point clouds, 3-dimensional position features and d-dimensional semantic features are split to obtain point cloud features with k2 point clouds, 3-dimensional position features and 0-dimensional semantic features and point cloud features with k2 point clouds, 0-dimensional position features and d-dimensional semantic features;

[0092] Perform position encoding processing on the point cloud features with k2 point clouds, 3-dimensional position features, and 0-dimensional semantic features to obtain position encoding features with k2 point clouds and d-dimensional position features;

[0093] The position coding features with k2 point clouds and d-dimensional position features are concatenated with the position coding features with k1 point clouds and d-dimensional position features, and then subjected to attention pooling to obtain enhanced position coding features.

[0094] The point cloud features with k1 points, 0-dimensional position features, and d-dimensional semantic features are concatenated with the point cloud features with k2 points, 0-dimensional position features, and d-dimensional semantic features, and then subjected to attention pooling processing to obtain enhanced point cloud features;

[0095] The enhanced position encoding features and the enhanced point cloud features are concatenated to obtain the target semantic features.

[0096] Among them, k1 and k2 represent the number of point clouds. k1 and k2 are not equal. The number of point clouds is k1 and the position feature is the d-dimensional position encoding feature. The mathematical expression is:

[0097]

[0098] in, Indicates that the number of point clouds is k1, and the position feature is the d-dimensional position encoding feature, Represents the coordinates of the central point cloud in the local area, Indicates point cloud features with k1 number of points, 3-dimensional position features, and 0-dimensional semantic features. The location information of the k1 adjacent point clouds is given by Given, symbol Indicates feature splicing processing, represents the Euclidean distance, The representation dimension transformation process is used to restore the encoded features to the feature dimensions of semantic information, thereby facilitating subsequent steps.

[0099] Furthermore, when encoding the geometric information in the local area, the present invention performs two KNN processes with different k values. When k1 and k2 are not equal, the present invention uses The present invention performs two different KNN operations on the original point cloud, which can enhance the geometric position information expression ability of the original point cloud and prevent some fuzzy segmentation points from interfering with local information. The mathematical expression of the enhanced position coding feature is:

[0100] ))

[0101] in, represents the enhanced position encoding feature, Indicates dimension transformation processing, represents attention calculation, It indicates that the number of point clouds is k2 and the position feature is a d-dimensional position encoding feature.

[0102] Specifically, ATT The principle is: ATTIt is widely used in the dilation feature extraction layer, and the subsequent calculation and interpretation are based on the assumption of local features, that is, the center point cloud is the point cloud The features of k adjacent point clouds. Usually, not all point cloud elements are the most important in a local area. More critical point cloud features should receive more attention. The attention pool is based on this theory and assigns an attention score to each point cloud feature. The higher the attention score, the more important the point cloud feature. Therefore, the present invention first calculates the attention scores of different point cloud features. The mathematical expression is:

[0103]

[0104] in, represents the attention score, represents the original features, Indicates dimension transformation processing, express Activation function.

[0105] Attention score Multiply by the original feature To generate attention features , then, the symmetric function SUM is used to aggregate features, thereby focusing on the entire local area. The mathematical expression is:

[0106]

[0107] in, Indicates the number of points in the point cloud.

[0108] After processing the geometric position information of the original point cloud, it is also necessary to locally expand and enhance other features of the original point cloud. The method adopted by the present invention is to first converge the features after two KNN operations, and then use Connect the two results. Since the method of the present invention emphasizes the aggregation of features in the expansion area, it is crucial to effectively utilize features of different scales. , so that these features can represent local areas while effectively collaborating to preserve feature details. However, it is inappropriate to directly use the output results as the final representation of the local structure at this stage, because the main goal of this invention is to enhance the generalization ability of the network from the point cloud data itself. Since features of different scales have been integrated in the local structure, it is chosen to combine different scales to improve network performance, so as to further identify adjacent point clouds in the high-dimensional feature space and obtain enhanced point cloud features. The mathematical expression is:

[0109]

[0110] in, Indicates that the number of point clouds is k1, the position feature is 0-dimensional, and the semantic feature is d-dimensional point cloud feature. It indicates that the number of point clouds is k2, the position feature is 0-dimensional, and the semantic feature is d-dimensional point cloud feature.

[0111] Furthermore, the target semantic features are obtained The mathematical expression is:

[0112]

[0113] For further information, see Figure 3 , which represents the structural diagram of the expanded feature extraction layer provided by the present invention, and the point cloud set is represented as: ,in, Indicates the number of points in the point cloud, each point cloud is defined as ,in Represents geometric information XYZ, Represents semantic information, which includes but is not limited to color information RGB and normal vector. For the dataset selected for the experiment in this invention, each point cloud is represented as ,In the extended feature extraction layer, the KNN algorithm is ,firstly used to establish the local region, which mainly ,focuses on aggregating the local features within the expanded region. ,Therefore, the present invention selects two different scales k1 and k2 for ,expanded feature extraction, where k1 ≠ k

[0114] In summary, the ERFA module of the present invention maximizes local feature representation while ensuring feature generalization performance. Specifically, ERFA focuses on feature aggregation in regions of two different scales. The core of ERFA is the expansion of the feature extraction layer and the addition of residual connections to address network degradation.

[0115] Step S3: query and interpolate each feature map with different dimensions respectively, obtain the semantic features corresponding to each feature map, and then perform feature splicing to generate the target semantic features.

[0116] Specifically, in the weakly supervised semantic segmentation task of point clouds, due to the limited number of labeled points, Figure 2 The semantic query module shown in the middle part is designed to efficiently utilize as many relevant features as possible. To query the center, search for nearby points based on Euclidean distance The semantic features of and The semantic feature combination of is used as the training signal input into the model.

[0117] The semantic query module's workflow consists of two steps: first, searching for neighboring features of the query point; second, performing feature interpolation between the query point and its neighboring points. After these two steps are completed, the semantic features corresponding to each feature map are obtained and then concatenated to generate the target semantic features.

[0118] Step S4: The preset initial prediction network model is trained according to the preset multi-scale classification strategy and local contrast regularization strategy to obtain the target prediction network model.

[0119] Specifically, in large-scale 3D point cloud scenes, the local features of the point cloud originate from the points and their surrounding areas, and these features often exhibit a high degree of semantic consistency. This semantic consistency indicates that points within a local area have similar semantic categories. For example, on continuous surfaces such as the walls of a building or the ground, adjacent points generally belong to the same category. Therefore, the accuracy of the semantic segmentation model depends largely on whether it can effectively capture and utilize these local features. If the model can accurately identify and utilize these local features, it can better understand the semantic structure of the scene, thereby improving segmentation accuracy.

[0120] In order to improve the segmentation performance, the present invention designs a local contrast regularization strategy to optimize the representation of local point cloud features. The core idea of ​​the local contrast regularization strategy is to enhance the model's perception of local features through contrast learning. Specifically, the strategy constructs positive sample pairs and negative sample pairs so that points in the same local area are closer in the feature space, while points in different local areas are pulled apart. This method can not only enhance the semantic consistency of local features, but also improve the model's sensitivity to local details. In addition, this local contrast regularization strategy can also be combined with multi-scale feature extraction to further improve the performance of the model.

[0121] Multi-scale feature extraction can capture multi-level information from local to global, while local contrast regularization focuses on enhancing the representation of local features. The two complement each other. In this way, the model can better identify and distinguish points of different semantic categories in complex 3D scenes, thereby achieving higher accuracy in large-scale point cloud semantic segmentation tasks. Specifically, for each 3D query point in the original point cloud and its corresponding point in the enhanced point cloud, the present invention introduces the coordinates XYZ of the corresponding point in the prediction space, and searches for the k nearest augmented neighbors through Euclidean distance, and calculates the average of the initial prediction probabilities of these augmented neighbors. .

[0122] To encourage The predicted value is closer to the average of the initial predicted probabilities , while being far away from the average prediction value of other neighbors, the present invention designs a local contrast regularization strategy loss function with the following formula.

[0123]

[0124]

[0125] in, represents the local contrast regularization strategy loss function, represents the number of points in the point cloud, represents the hyperparameters set, represents the number of nearest neighbor points, Represents the original point cloud Through the function The predicted value obtained after processing, Indicates parameters, Represents a query point The predicted value of represents the mean of the predicted values, represents the indicator function, when When hour, ,when When hour, , this indicator function is used to distinguish whether the current point and the nearest neighbor point of the current point are the same.

[0126] Furthermore, the foreground and background in large-scale point clouds often exhibit significant scale differences. Multi-scale classification strategies can effectively integrate global and local information, more accurately capturing the characteristic differences between the two. To further improve segmentation accuracy, the concept of multi-scale classification was introduced to enhance the model's adaptability to different scenes and objects.

[0127] The multi-scale classification strategy can extract the global structural features of the background at a coarse scale, while capturing the local detail features of foreground objects at a fine scale. This multi-scale feature fusion method enables the model to better understand the overall semantic structure of the scene while paying attention to local details, thereby improving the accuracy of segmentation. Specifically, the original point cloud is first downsampled multiple times and features are aggregated to extract multi-scale features; then, the local features of the query point are extracted, and the classification task is performed at each scale. The specific approach is: the features of each scale are input into a two-layer fully connected network for classification, and then the classification results are integrated and incorporated into the segmentation loss function. By fusing multi-scale features, the model can fully capture the global overview and fine local details, thereby optimizing the entire training process. The mathematical expression of the loss function of the multi-scale classification strategy is:

[0128]

[0129] in, represents the multi-scale classification strategy loss function, is the total number of labeled classes, represents the number of points in the point cloud, is the total number of layers, For The true value of the class, For Class The predicted value of the layer.

[0130] Step S5: Input the target semantic features as training signals into the target prediction network model for processing to generate classification results.

[0131] Specifically, the target semantic features include the semantic features corresponding to the query point cloud and the semantic features corresponding to the neighboring point clouds of the query point cloud. When the target semantic features are input as training signals into the target prediction network model for processing, the classification results are obtained.

[0132] In summary, the present invention uses the ERFA module as its core to achieve the expansion and enhancement of local region features. It also combines multi-scale classification and local contrast regularization strategies to fully utilize the semantic information in sparsely labeled data to generate classification results. The workflow of the overall structure diagram of multi-scale feature extraction and classification provided by the present invention includes the following steps:

[0133] First, the input point cloud data is preprocessed, and the features of the local area of ​​the point cloud are expanded and enhanced through the ERFA module to extract more representative local features;

[0134] Secondly, feature aggregation is performed on the original point cloud data and the enhanced data respectively, and a random downsampling strategy is adopted to reduce the computational complexity and retain key information.

[0135] Then, based on the randomly selected point cloud coordinates and multi-scale features, the semantic features of the local area are captured and input into the prediction network to generate the initial classification results;

[0136] Finally, through further feature fusion and optimization, the final semantic segmentation classification results are obtained. During the model training phase, a local contrast regularization strategy is used to iteratively optimize the model. Through multi-scale classification methods, the model is able to extract a wide range of detailed structural features, significantly enhancing its generalization ability across different scenarios. In addition, the local contrast regularization strategy fully exploits the local feature information implicit in the weakly supervised labels, further improving the model's perception of local details.

[0137] Through this multi-strategy fusion approach, the present invention performs well in the task of 3D point cloud semantic segmentation in large-scale scenes, and can achieve high-precision semantic segmentation effects under conditions of limited annotated data, providing an effective solution for point cloud processing in weakly supervised scenarios.

[0138] Example 2

[0139] See also Figure 5 , which represents the internal structure of the weakly supervised point cloud semantic segmentation system for multi-scale feature extraction and classification provided by the present invention, including:

[0140] The preprocessing module 100 is used to obtain and preprocess the original point cloud data to generate target point cloud data, where the target point cloud data consists of the number, position features, and semantic features of the point cloud;

[0141] ERFA module 200 is used to perform feature extraction and feature aggregation processing on the target point cloud data to generate multiple feature maps with different dimensions;

[0142] The target semantic feature generation module 300 is used to query and interpolate the feature maps of different dimensions, obtain the semantic features corresponding to each feature map, and then perform feature concatenation to generate target semantic features. The target semantic features include the semantic features corresponding to the query point cloud and the semantic features corresponding to the neighboring point clouds of the query point cloud.

[0143] The target preset network model generation module 400 is used to train the preset initial prediction network model according to the preset multi-scale classification strategy and local contrast regularization strategy to obtain the target prediction network model;

[0144] The classification module 500 is used to input the target semantic features as training signals into the target prediction network model for processing to generate classification results.

[0145] The beneficial effect of the weakly supervised point cloud semantic segmentation system with multi-scale feature extraction and classification provided by the present invention is that:

[0146] First, through the ERFA module 200, the features of the local area of ​​the point cloud are expanded and enhanced to extract more representative local features. The ERFA module 200 can ensure the feature generalization performance while maximizing the feature representation of the local area. The ERFA module 200 uses the internal expanded feature extraction layer to consider the aggregation of local features of different scales and the information in high-dimensional and low-dimensional feature spaces, and solves the network degradation problem through feature splicing. The ERFA module 200 has stronger generalization performance and can use marker points to fully generalize features in weakly supervised semantic segmentation tasks to enhance the performance of the entire network.

[0147] Second, through the target preset network model generation module 400, it is used to train the preset initial prediction network model according to the preset multi-scale classification strategy and local contrast regularization strategy, which can obtain more stable features from the disordered point cloud data, making the classification results generated by the prediction network model more accurate.

[0148] The weakly supervised point cloud semantic segmentation system for multi-scale feature extraction and classification in the embodiments of the present application can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, the mobile electronic device can be a mobile phone, tablet computer, laptop computer, PDA, in-vehicle electronic device, wearable device, ultra-mobile personal computer (UMPC), netbook or personal digital assistant (PDA), etc. The non-mobile electronic device can be a server, network attached storage (NAS), personal computer (PC), television (TV), ATM or self-service machine, etc., and the embodiments of the present application do not specifically limit this.

[0149] The weakly supervised point cloud semantic segmentation system for multi-scale feature extraction and classification in the embodiments of the present application can be a device having an operating system. The operating system can be an Android operating system, an iOS operating system, or other possible operating systems, which are not specifically limited in the embodiments of the present application.

[0150] The weakly supervised point cloud semantic segmentation system provided by the embodiment of the present application can achieve Figures 1 to 4 To avoid repetition, the various processes implemented by the weakly supervised point cloud semantic segmentation system for multi-scale feature extraction and classification in the method embodiment are not described here.

[0151] Optionally, an embodiment of the present application also provides an electronic device, including a processor, a memory, and a program or instruction stored in the memory and executable on the processor. When the program or instruction is executed by the processor, each process of the embodiment of the weakly supervised point cloud semantic segmentation method for multi-scale feature extraction and classification is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be described here.

[0152] An embodiment of the present application also provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, each process of the embodiment of the weakly supervised point cloud semantic segmentation method for multi-scale feature extraction and classification is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0153] The processor is the processor in the electronic device described in the above embodiment. The readable storage medium includes a computer-readable storage medium, such as a computer read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk.

[0154] It should be noted that, in the present invention, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in the opposite order according to the functions involved. For example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0155] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a more preferred embodiment. Based on this understanding, the technical solution of this application, or the part that contributes to the existing technology, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of this application.

[0156] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of this application, ordinary technicians in this field can also make many forms without departing from the purpose of this application and the scope of protection of the claims, all of which are within the protection of this application.

Claims

1. A weakly supervised point cloud semantic segmentation method based on multi-scale feature extraction and classification, characterized by: include: Acquire and preprocess raw point cloud data to generate target point cloud data, wherein the target point cloud data consists of the number, position features, and semantic features of the point cloud; Performing feature extraction and feature aggregation processing on the target point cloud data according to a preset ERFA module to generate multiple feature maps with different dimensions, wherein the preset ERFA module specifically includes: a first shared MLP layer, a second shared MLP layer, a third shared MLP layer, an expanded feature extraction layer, and an activation function layer, wherein the second shared MLP layer, the expanded feature extraction layer, the third shared MLP layer, and the activation function layer are sequentially connected in series from left to right, and the first shared MLP layer is connected in parallel with the second shared MLP layer, the expanded feature extraction layer, the third shared MLP layer, and the activation function layer; Query and interpolate each feature map with different dimensions respectively, obtain the semantic features corresponding to each feature map and then perform feature splicing processing to generate target semantic features, wherein the expansion feature extraction layer is used to obtain aggregate features, and the specific steps include: processing the input point cloud features to obtain point cloud features with a point cloud quantity of 1, a position feature of 3 dimensions, and a semantic feature of d dimensions; splitting the point cloud features with a point cloud quantity of 1, a position feature of 3 dimensions, and a semantic feature of d dimensions to obtain point cloud features with a point cloud quantity of 1, a position feature of 3 dimensions, and a semantic feature of d dimensions, and obtaining point cloud features with a point cloud quantity of 1, a position feature of 3 dimensions, and a semantic feature of 0 dimensions, as well as a point cloud quantity of 1, a position feature of 0 dimensions, and a semantic feature of d dimensions. The feature is a d-dimensional point cloud feature; the first KNN processing is performed on the point cloud feature with a point cloud quantity of 1, a position feature of 3 dimensions, and a semantic feature of 0 dimension, and the point cloud feature with a point cloud quantity of 1, a position feature of 0 dimension, and a semantic feature of d dimension, respectively, and the point cloud feature with a point cloud quantity of k1, a position feature of 3 dimensions, and a semantic feature of 0 dimension and the point cloud feature with a point cloud quantity of k1, a position feature of 0 dimension, and a semantic feature of d dimension are obtained; the point cloud feature with a point cloud quantity of k1, a position feature of 3 dimensions, and a semantic feature of 0 dimension is performed position coding processing to obtain a position coding feature with a point cloud quantity of k1 and a position feature of d dimension; the point cloud feature with a point cloud quantity of 1. The point cloud features with 3-dimensional position features and d-dimensional semantic features are subjected to the second KNN processing to obtain point cloud features with k2 point clouds, 3-dimensional position features and d-dimensional semantic features, wherein the focus scales are different during the first KNN processing and the second KNN processing; the point cloud features with k2 point clouds, 3-dimensional position features and d-dimensional semantic features are split to obtain point cloud features with k2 point clouds, 3-dimensional position features and 0-dimensional semantic features and point cloud features with k2 point clouds, 0-dimensional position features and d-dimensional semantic features; the point cloud features with k2 point clouds, 3-dimensional position features and 0-dimensional semantic features are subjected to the second KNN processing to obtain point cloud features with k2 point clouds, 3-dimensional position features and 0-dimensional semantic features. Position coding processing is performed to obtain position coding features with a point cloud quantity of k2 and a position feature of d dimensions; the position coding features with a point cloud quantity of k2 and a position feature of d dimensions are subjected to feature splicing processing and then attention pooling processing are performed to obtain enhanced position coding features; the point cloud features with a point cloud quantity of k1, a position feature of 0 dimensions and a semantic feature of d dimensions are subjected to feature splicing processing and then attention pooling processing are performed to obtain enhanced point cloud features, wherein the enhanced point cloud features The mathematical expression is: ； in, Indicates that the number of point clouds is k1, the position feature is 0-dimensional, and the semantic feature is d-dimensional point cloud feature. Indicates point cloud features with k2 number of points, 0-dimensional position features, and d-dimensional semantic features; the enhanced position coding features and the enhanced point cloud features are subjected to feature splicing processing to obtain aggregate features, wherein the aggregate features The mathematical expression is: ； in, Representing the enhanced position encoding features; training the preset initial prediction network model according to the preset multi-scale classification strategy and local contrast regularization strategy to obtain the target prediction network model; The target semantic features are input as training signals into the target prediction network model for processing to generate classification results.

2. The weakly supervised point cloud semantic segmentation method based on multi-scale feature extraction and classification according to claim 1 is characterized in that: The specific process of performing feature extraction and feature aggregation processing on the target point cloud data according to the preset ERFA module to generate multiple feature maps with different dimensions includes: Performing feature extraction and feature aggregation processing on the target point cloud data according to the first ERFA module to generate a first feature map; performing feature extraction and feature aggregation processing on the first feature map according to the second ERFA module to generate a second feature map; performing feature extraction and feature aggregation processing on the second feature map according to the third ERFA module to generate a third feature map; performing feature extraction and feature aggregation processing on the third feature map according to the fourth ERFA module to generate a fourth feature map; Among them, the semantic feature dimensions of the first feature map, the second feature map, the third feature map, and the fourth feature map are the same, the position feature dimension of the fourth feature map is higher than that of the third feature map, the position feature dimension of the third feature map is higher than that of the second feature map, the position feature dimension of the second feature map is higher than that of the first feature map, the number of point clouds of the fourth feature map is lower than that of the third feature map, the number of point clouds of the third feature map is lower than that of the second feature map, and the number of point clouds of the second feature map is lower than that of the first feature map.

3. The weakly supervised point cloud semantic segmentation method based on multi-scale feature extraction and classification according to claim 2 is characterized in that: The specific steps of performing feature extraction and feature aggregation processing on the target point cloud data according to the first ERFA module to generate a first feature map include: Performing dimension transformation processing on the target point cloud data according to the first shared MLP layer and the second shared MLP layer, respectively, to correspondingly obtain a first semantic feature after dimension transformation and a second semantic feature after dimension transformation; Performing position encoding processing on the second semantic feature after the dimension transformation to generate an input point cloud feature; Performing expansion feature extraction processing on the input point cloud features to generate first aggregated features; Performing dimension transformation on the first aggregated feature according to the third shared MLP layer and performing feature splicing on the first semantic feature after the dimension transformation to obtain a spliced ​​feature; The concatenated features are processed according to an activation function to generate a first feature map.

4. The weakly supervised point cloud semantic segmentation method based on multi-scale feature extraction and classification according to claim 1 is characterized in that: The local contrast regularization strategy includes: in, represents the local contrast regularization strategy loss function, represents the number of points in the point cloud, represents the hyperparameters set, represents the number of nearest neighbor points, Represents the original point cloud Through the function The predicted value obtained after processing, Indicates parameters, Represents a query point The predicted value of represents the mean of the predicted values, Represents the indicator function.

5. The weakly supervised point cloud semantic segmentation method based on multi-scale feature extraction and classification according to claim 1 is characterized in that: The multi-scale classification strategy includes: in, represents the multi-scale classification strategy loss function, is the total number of labeled classes, represents the number of points in the point cloud, is the total number of layers, For The true value of the class, For Class The predicted value of the layer.

6. A weakly supervised point cloud semantic segmentation system with multi-scale feature extraction and classification, characterized by: include: A preprocessing module is used to acquire and preprocess the original point cloud data to generate target point cloud data, wherein the target point cloud data is composed of the number, position features and semantic features of the point cloud; An ERFA module is configured to perform feature extraction and feature aggregation processing on the target point cloud data to generate multiple feature maps of different dimensions; wherein the ERFA module specifically includes: a first shared MLP layer, a second shared MLP layer, a third shared MLP layer, an expanded feature extraction layer, and an activation function layer, wherein the second shared MLP layer, the expanded feature extraction layer, the third shared MLP layer, and the activation function layer are sequentially connected in series from left to right, and the first shared MLP layer is connected in parallel with the second shared MLP layer, the expanded feature extraction layer, the third shared MLP layer, and the activation function layer; The target semantic feature generation module is used to query and interpolate each feature map with different dimensions respectively, and perform feature splicing processing after obtaining the semantic features corresponding to each feature map to generate target semantic features. The target semantic features include the semantic features corresponding to the query point cloud and the semantic features corresponding to the neighboring point clouds of the query point cloud; wherein, the expansion feature extraction layer is used to obtain aggregate features, and the specific process of obtaining the aggregate features includes: processing the input point cloud features to obtain point cloud features with a point cloud quantity of 1, a position feature of 3 dimensions, and a semantic feature of d dimensions; splitting the point cloud features with a point cloud quantity of 1, a position feature of 3 dimensions, and a semantic feature of d dimensions to obtain Take the point cloud features with a point cloud quantity of 1, a position feature of 3 dimensions, and a semantic feature of 0 dimension, and the point cloud features with a point cloud quantity of 1, a position feature of 0 dimension, and a semantic feature of d dimension; respectively perform the first KNN processing on the point cloud features with a point cloud quantity of 1, a position feature of 3 dimensions, and a semantic feature of 0 dimension, and the point cloud features with a point cloud quantity of 1, a position feature of 0 dimension, and a semantic feature of d dimension, and correspondingly obtain the point cloud features with a point cloud quantity of k1, a position feature of 3 dimensions, and a semantic feature of 0 dimension, and the point cloud features with a point cloud quantity of k1, a position feature of 0 dimension, and a semantic feature of d dimension; perform position encoding processing on the point cloud features with a point cloud quantity of k1, a position feature of 3 dimensions, and a semantic feature of 0 dimension to obtain Take the position coding feature with k1 point cloud number and d-dimensional position feature; perform the second KNN processing on the point cloud feature with 1 point cloud number, 3-dimensional position feature and d-dimensional semantic feature to obtain the point cloud feature with k2 point cloud number, 3-dimensional position feature and d-dimensional semantic feature, wherein the focus scale is different during the first KNN processing and the second KNN processing; split the point cloud feature with k2 point cloud number, 3-dimensional position feature and d-dimensional semantic feature to obtain the point cloud feature with k2 point cloud number, 3-dimensional position feature and 0-dimensional semantic feature and the point cloud feature with k2 point cloud number, 0-dimensional position feature and d-dimensional semantic feature; perform the second KNN processing on the point cloud feature with k2 point cloud number, 3-dimensional position feature and d-dimensional semantic feature The position coding process is performed on the point cloud features with k2 point clouds and d-dimensional position features, and the position coding features with k1 point clouds and d-dimensional position features are processed by feature splicing, and then the attention pooling process is performed to obtain enhanced position coding features; the point cloud features with k1 point clouds, 0-dimensional position features and d-dimensional semantic features are processed by feature splicing, and then the attention pooling process is performed to obtain enhanced point cloud features, wherein the enhanced point cloud features are processed. The mathematical expression is: ； in, Indicates that the number of point clouds is k1, the position feature is 0-dimensional, and the semantic feature is d-dimensional point cloud feature. Indicates point cloud features with k2 number of points, 0-dimensional position features, and d-dimensional semantic features; the enhanced position coding features and the enhanced point cloud features are subjected to feature splicing processing to obtain aggregate features, wherein the aggregate features The mathematical expression is: ； in, Represents the enhanced position encoding feature; A target preset network model generation module is used to train the preset initial prediction network model according to the preset multi-scale classification strategy and local contrast regularization strategy to obtain the target prediction network model; The classification module is used to input the target semantic features as training signals into the target prediction network model for processing to generate classification results.

7. An electronic device, characterized in that: It includes a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein when the program or instruction is executed by the processor, the steps of the weakly supervised point cloud semantic segmentation method for multi-scale feature extraction and classification are implemented as described in any one of claims 1 to 5.

8. A readable storage medium, characterized in that: The readable storage medium stores a program or instruction, and when the program or instruction is executed by the processor, the steps of the weakly supervised point cloud semantic segmentation method for multi-scale feature extraction and classification as described in any one of claims 1 to 5 are implemented.

Images