A small sample point cloud semantic segmentation method based on difference enhancement and related equipment

Abstract

The application discloses a small sample point cloud semantic segmentation method based on difference enhancement and related equipment, and the method comprises the following steps: obtaining initial prototypes of various categories according to support set features and corresponding support set masks; calculating the similarity between the initial prototypes and query set features to obtain the probability distribution of the query set features belonging to foreground and background categories; taking the inverse of the foreground distribution according to foreground and background difference enhancement, combining the current background distribution to obtain accurate background probability distribution, and integrating to obtain reliable probability distribution of the entire category; the query set features are subjected to transformation processing through a feature aggregation module to obtain aggregated query set features; the reliable probability distribution and the aggregated query set features are subjected to aggregation processing through a prototype correction module to obtain pseudo prototypes; the support set data is expanded according to the pseudo prototypes through the prototype correction module to obtain corrected prototypes, and the segmentation result of the point cloud is obtained according to the corrected prototypes. The application obtains better segmentation effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a small sample point cloud semantic segmentation method based on difference enhancement, a terminal and a computer readable storage medium. BACKGROUND

[0002] Small sample learning relies on a small amount of data and its mask, but when dealing with multiple foreground classes, the mask of these data can only indicate whether the data belongs to a specific class, and cannot indicate whether the data belongs to other foreground classes or background. In this case, the foreground of some classes is regarded as background, that is, background ambiguity occurs, so that the existing solutions including the prototype network are difficult to distinguish between background and foreground, inevitably leading to serious prototype bias and incorrect segmentation results. Although background ambiguity has a significant impact on the model, this problem has not been given enough attention.

[0003] The small sample point cloud semantic segmentation method refers to the condition that a small amount of labeled new class samples are given, and the model can perform semantic segmentation on the point cloud to obtain the point cloud belonging to the new class. In order to extract effective point cloud features, researchers strive to obtain features that can better represent class information. For example, the average value of class features is taken as a prototype, and the distance between point features and the prototype is used to segment the point cloud. However, this method does not take into account the complexity of the structure of the point cloud itself, and the segmentation accuracy is not satisfactory. For example, multiple prototypes are used to describe complex point cloud distribution, but the reasoning time is longer.

[0004] The prior art ignores the background ambiguity existing at the current data level. The goal of small sample learning is to predict the label of query data given some annotated support data, which usually includes data and corresponding labels grouped by class. However, the foreground of one class may be mistakenly regarded as the background of other classes, which is called background ambiguity.

[0005] Therefore, the prior art still needs to be improved and developed. SUMMARY

[0006] The main purpose of the present application is to provide a small sample point cloud semantic segmentation method based on difference enhancement, a terminal and a computer readable storage medium, which aims to solve the problem that the small sample point cloud semantic segmentation method in the prior art is prone to background ambiguity, resulting in low small sample point cloud semantic segmentation accuracy.

[0007] To achieve the above purpose, the present application provides a small sample point cloud semantic segmentation method based on difference enhancement, which comprises the following steps:

[0008] Obtain query set data and support set data, and perform feature alignment processing on the query set data and the support set data to obtain query set features and support set features;

[0009] obtaining initial prototypes of each class according to the support set features and corresponding support set masks, wherein the initial prototypes include prototypes of multiple classes;

[0010] calculating similarities between the initial prototypes and the query set features to obtain probability distributions of the query set features belonging to foreground and background classes;

[0011] taking inverse of the foreground distribution according to the foreground-background difference enhancement, combining the foreground distribution with the current background distribution to obtain an accurate background probability distribution, and integrating to obtain reliable probability distributions of the entire class;

[0012] transforming the query set features through a feature aggregation module to obtain aggregated query set features;

[0013] obtaining pseudo prototypes through aggregation processing of the reliable probability distributions and the aggregated query set features by a prototype correction module;

[0014] obtaining corrected prototypes by the prototype correction module according to the pseudo prototypes to expand the support set data, and obtaining segmentation results of the point cloud according to the corrected prototypes.

[0015] Optionally, the small sample point cloud semantic segmentation method based on difference enhancement, wherein the feature alignment processing according to the query set data and the support set data to obtain query set features and support set features specifically includes:

[0016] calculating distances between two features in the query set data and the support set data;

[0017] taking inverse of the distances and performing normalization processing to obtain corresponding weights;

[0018] obtaining query set features according to the query set data, multiplying the obtained weights with the query set features to obtain aligned support set features.

[0019] Optionally, the small sample point cloud semantic segmentation method based on difference enhancement, wherein the initial prototypes of each class are obtained according to the support set features and corresponding support set masks, specifically including:

[0020] obtaining initial prototypes P of each class by using aligned support set features F s and corresponding support set masks M:

[0021]

[0022] wherein the initial prototypes P include prototypes of multiple classes.

[0023] Optionally, the small sample point cloud semantic segmentation method based on difference enhancement, wherein the similarity of the initial prototype and the query set feature is calculated to obtain the probability distribution of the query set feature belonging to each of the foreground and background categories, and specifically includes:

[0024] The similarity of the initial prototype P and the query set feature F is calculated to obtain the probability distribution of the query set feature belonging to a specific category. q

[0025] For the i-th prototype p i , the probability of the query set feature belonging to category i is defined as:

[0026]

[0027] wherein, denotes the distance between the prototype and the query set feature; and channel denotes the number of feature channels.

[0028] Optionally, the small sample point cloud semantic segmentation method based on difference enhancement, wherein the foreground distribution is negated based on foreground-background difference enhancement, combined with the current background distribution to obtain an accurate background probability distribution, and reliable probability distributions of the entire category are integrated, and specifically includes:

[0029] The probability distribution of the query set feature belonging to each of the foreground and background categories is obtained based on foreground-background difference enhancement, the foreground distribution is negated based on foreground-background difference enhancement, and combined with the current background distribution to obtain an accurate background probability distribution

[0030]

[0031] wherein, Prob 0 is the background probability distribution, the foreground probability distribution Prob i is the i-th foreground probability distribution, and C is the number of foreground categories.

[0032] The reliable background probability distribution and the foreground probability distribution are integrated to obtain a reliable probability distribution of the entire category.

[0033]

[0034] Optionally, the small sample point cloud semantic segmentation method based on difference enhancement, wherein the query set feature is processed by a feature aggregation module to obtain an aggregated query set feature, and specifically includes:

[0035] The feature aggregation module based on geometric guidance selects points within the k-neighborhood of each geometric point, and for a point x i within X, the points within the k-NN range​ is represented as:

[0036]

[0037] wherein q is a distance between x i the kth nearest point;

[0038] The features obtained through the feature extractor have a one-to-one correspondence with the geometric points, and the serial number of each geometric point is the same as the serial number of the corresponding feature. According to the one-to-one correspondence between the geometric points and the features, the features corresponding to the nearest neighbor points are obtained from the features

[0039] The feature aggregation module based on geometric guidance obtains the weight of each point relative to the initial point according to the distance between the k-NN inner point and the initial point:

[0040]

[0041] wherein, denotes the distance between two points;

[0042] According to the weight, the feature aggregation module based on geometric guidance aggregates the local features, and increases the point feature information to retain the information of the point itself, to obtain the aggregated features. For a certain point feature f i , the corresponding feature after aggregation is defined as:

[0043]

[0044] wherein w represents a learnable parameter;

[0045] After the feature aggregation module based on geometric guidance, the query set feature F q is converted into the aggregated query set feature

[0046] Optionally, the small sample point cloud semantic segmentation method based on difference enhancement, wherein the aggregation processing of the reliable probability distribution and the aggregated query set feature through the prototype correction module obtains a pseudo prototype, specifically comprising:

[0047] According to the reliable probability distribution Prob reli of all categories based on the prototype correction module, the n features f top with the highest probability in each probability distribution are selected from the geometric aggregated query set feature . For category i, the selected feature f is defined as:

[0048]

[0049] wherein probi,j is the jth probability in the ith class, m is the nth largest probability in the ith class, based on the one-to-one correspondence between the feature and the probability, the serial number of each feature is the same as the serial number of the corresponding probability, and the feature f top corresponding probability prob top ;

[0050] The prototype correction module is aggregated based on a plurality of features and corresponding probabilities to obtain a pseudo prototype:

[0051]

[0052] Optionally, the small sample point cloud semantic segmentation method based on difference enhancement, wherein the support set data is expanded according to the pseudo prototype by the prototype correction module to obtain a corrected prototype, and a segmentation result of the point cloud is obtained according to the corrected prototype, specifically including:

[0053] The support set data is expanded by the prototype correction module through the pseudo prototype to obtain a corrected prototype:

[0054]

[0055] Wherein, N is the number of data of the prototype corresponding category in the support set data;

[0056] After the corrected prototype is subjected to distribution normalization processing and similarity calculation, a segmentation result of the point cloud is obtained.

[0057] In addition, in order to achieve the above-mentioned purpose, the application further provides a terminal, wherein the terminal comprises a memory, a processor and a small sample point cloud semantic segmentation program based on difference enhancement stored on the memory and executable on the processor, and the small sample point cloud semantic segmentation program based on difference enhancement realizes the steps of the small sample point cloud semantic segmentation method based on difference enhancement when the processor is executed.

[0058] In addition, in order to achieve the above-mentioned purpose, the application further provides a computer readable storage medium, wherein the computer readable storage medium stores a small sample point cloud semantic segmentation program based on difference enhancement, and the small sample point cloud semantic segmentation program based on difference enhancement realizes the steps of the small sample point cloud semantic segmentation method based on difference enhancement when the processor is executed.

[0059] In the present application, query set data and support set data are obtained, feature alignment processing is performed according to the query set data and the support set data to obtain query set features and support set features, initial prototypes of various categories are obtained according to the support set features and corresponding support set masks, the initial prototypes contain prototypes of multiple categories, the similarity of the initial prototypes and the query set features is calculated to obtain the probability distribution of the query set features belonging to foreground and background categories, the foreground distribution is negated according to foreground-background difference enhancement, combined with the current background distribution to obtain accurate background probability distribution, and reliable probability distribution of the entire category is integrated, the query set features are converted by a feature aggregation module to obtain aggregated query set features, the reliable probability distribution and the aggregated query set features are aggregated by a prototype correction module to obtain pseudo prototypes, the support set data is expanded according to the pseudo prototypes by the prototype correction module to obtain corrected prototypes, and the segmentation result of the point cloud is obtained according to the corrected prototypes. The present application overcomes background ambiguity through progressive foreground-background enhancement, removes outliers through a feature aggregation module under geometric guidance, and corrects prototypes using a prototype correction module, thereby obtaining better segmentation results. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is a flowchart of a preferred embodiment of the present application based on difference-enhanced small sample point cloud semantic segmentation method;

[0061] Figure 2 is a principle diagram of the entire semantic style process in the preferred embodiment of the present application based on difference-enhanced small sample point cloud semantic segmentation method;

[0062] Figure 3 is a running environment diagram of the preferred embodiment of the terminal of the present application. DETAILED DESCRIPTION

[0063] To make the purpose, technical scheme and advantages of the present application clearer and more explicit, the present application is further described in detail below with reference to the drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0064] To solve the existing background ambiguity problem, the present application first uses a foreground-background difference enhancement (BFDE) strategy, uses reliable foreground to eliminate the interference information of the background, thereby obtaining reliable foreground and background probability distribution, then extracts reliable features according to the reliable probability distribution, and aggregates the features into pseudo prototypes, and uses the pseudo prototypes to expand the features to correct the prototypes.

[0065] The small sample point cloud semantic segmentation method based on difference enhancement provided by the preferred embodiment of the present application comprises the following steps: Figure 1 and Figure 2 The small sample point cloud semantic segmentation method based on difference enhancement comprises the following steps:

[0066] In step S10, query set data and support set data are obtained, and feature alignment processing is performed on the query set data and the support set data to obtain query set features and support set features.

[0067] Specifically, as shown in Figure 2 , the feature extractor extracts features using a dynamic graph convolutional neural network, and then performs feature alignment operation, which aligns the features by reducing the distance between the support set and the query set features, and the specific steps include:

[0068] (1) Calculate the distance between the features in the query set data and the support set data;

[0069] (2) Take the inverse of the distance and perform normalization processing to obtain the corresponding weight;

[0070] (3) Multiply the obtained weight with the query set features to obtain the aligned support set features.

[0071] In step S20, initial prototypes of each class are obtained according to the support set features and the corresponding support set masks, and the initial prototypes include prototypes of multiple classes.

[0072] Specifically, as shown in Figure 2 , the foreground-background difference enhancement first obtains initial prototypes P of each class using the aligned support set features F s and the corresponding support set masks M:

[0073]

[0074] The initial prototypes P include prototypes of multiple classes.

[0075] In step S30, the similarity between the initial prototypes and the query set features is calculated to obtain the probability distribution of the query set features belonging to each class of foreground and background.

[0076] Specifically, as shown in Figure 2 , the foreground-background difference enhancement calculates the similarity between the initial prototypes P and the query set features F q , thereby obtaining the probability distribution of the query set features belonging to a specific class.

[0077] For the i-th class of prototype p i , the probability of the query set features belonging to class i is defined as:

[0078]

[0079] wherein, denotes the distance between the prototype and the query set feature; channel denotes the number of feature channels.

[0080] Step S40, according to the foreground-background difference enhancement, the foreground distribution is negated, combined with the current background distribution to obtain an accurate background probability distribution, and the reliable probability distribution of the entire category is integrated.

[0081] Specifically, as shown in Figure 2 , through the above operation, the foreground-background difference enhancement obtains the probability distribution of the query set feature belonging to each category of foreground and background. Due to the background ambiguity problem, the probability distribution of the current background category is inaccurate. Considering that the foreground distribution is accurate, the foreground distribution is negated based on the foreground-background difference enhancement, and combined with the current background distribution to obtain an accurate background probability distribution

[0082]

[0083] wherein, Prob 0 is the background probability distribution, the foreground probability distribution Prob i is the i-th foreground probability distribution, and C is the number of foreground categories.

[0084] By integrating the reliable background probability distribution and the foreground probability distribution, the reliable probability distribution of the entire category can be obtained:

[0085]

[0086] Step S50, the query set feature is processed by a feature aggregation module to obtain an aggregated query set feature.

[0087] Specifically, as shown in Figure 2 , the Geometric-guided Feature Aggregation Module (GFAM) uses geometric information to select points within the k-nearest neighbors (k-NN) range of each geometric point. For a point x i in X, the points in the k-NN range of x are denoted as:

[0088]

[0089] wherein, q is the point closest to x i .

[0090] Note that the features obtained by the feature extractor have a one-to-one correspondence with geometric points; the index of each geometric point is the same as the index of its corresponding feature. Based on this one-to-one correspondence between geometric points and features, the features corresponding to the nearest neighbors can be obtained from the features.

[0091] Then, the geometry-guided feature aggregation module obtains the weights of each point relative to the initial point based on the distance between the points in the k-NN and the initial point:

[0092]

[0093] in, It represents the distance between two points.

[0094] Based on weights, the geometry-guided feature aggregation module aggregates local features and adds point feature information while retaining the information of the points themselves, thus obtaining the aggregated features. For a certain point feature f i The corresponding features after aggregation Defined as:

[0095]

[0096] Here, w represents a learnable parameter.

[0097] The feature aggregation module guided by geometry will aggregate the query set features F q Transformed into aggregated query set features

[0098] Step S60: Based on the reliable probability distribution and the aggregated query set features, the prototype correction module performs aggregation processing to obtain a pseudo prototype.

[0099] Specifically, such as Figure 2 As shown, the Prototype Refinement Module (PRM) is based on a reliable probability distribution of all categories. reli Features of the query set after geometric aggregation Select the n features f with the highest probability from various probability distributions. top For category i, the selected features Defined as:

[0100]

[0101] Among them, prob i,j This refers to the j-th probability in the i-th class, where m is the n-th largest probability in the i-th class. Based on the one-to-one correspondence between features and probabilities—the index of each feature is the same as the index of its corresponding probability—feature f can be obtained from the index. topThe corresponding probability prob top .

[0102] Then, the plurality of features and the corresponding probabilities are aggregated based on the prototype correction module to obtain a pseudo prototype:

[0103]

[0104] In step S70, the support set data is expanded according to the pseudo prototype through the prototype correction module to obtain a corrected prototype, and a segmentation result of the point cloud is obtained according to the corrected prototype.

[0105] Specifically, as shown in Figure 2 , the support set data is expanded based on the prototype correction module through the pseudo prototype to obtain a corrected prototype:

[0106]

[0107] Wherein, N is the number of data of the prototype corresponding category in the support set data.

[0108] After the corrected prototype is subjected to distribution normalization processing and similarity calculation, a segmentation result of the point cloud is obtained.

[0109] Wherein, the present application adjusts the feature distribution through a self-attention module, and the distribution is normalized:

[0110] (1). The aggregated query set features are spliced with the corrected prototype to obtain expanded features;

[0111] (2). A self-attention module is used to adjust the distribution normalization of the expanded features;

[0112] (3). The adjusted expanded features are segmented to obtain corresponding distribution normalized features and distribution normalized prototypes.

[0113] Wherein, the present application obtains the similarity by calculating the distance between the features and the prototype, and the similarity calculation:

[0114] (1). The distance between the prototype and the feature is calculated;

[0115] (2) The distance is taken as the opposite and normalized to obtain the similarity of the feature and each type of prototype. Note that, in the test, the category of the feature most similar to the prototype is selected as the prediction result, so that the segmentation result of the point cloud is obtained.

[0116] In addition, after the similarity calculation, cross entropy loss processing will be performed according to the query set label, the similarity of the feature and the prototype will be regarded as the probability that the feature belongs to these categories, and the corresponding cross entropy loss will be calculated according to the query set label and the cross entropy function, as the error of the current result.

[0117] In summary, the application overcomes background ambiguity by progressive foreground-background enhancement, removes outliers by a feature aggregation module under geometric guidance, and corrects prototypes by a prototype correction module, thereby obtaining better segmentation results.

[0118] The application has improved semantic segmentation accuracy on several commonly used data sets through multiple group experiments compared with other small sample point cloud semantic segmentation methods.

[0119] Further, as shown in Figure 3 Based on the above small sample point cloud semantic segmentation method and system based on difference enhancement, the application further provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 3 Only part of the components of the terminal are shown, but it should be understood that all the shown components are not required to be implemented, and more or less components can be alternatively implemented.

[0120] The memory 20 can be an internal storage unit of the terminal in some embodiments, such as a hard disk or a memory of the terminal. The memory 20 can also be an external storage device of the terminal in other embodiments, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal. Further, the memory 20 can include both the internal storage unit and the external storage device of the terminal. The memory 20 is used to store application software and various data installed on the terminal, such as program codes of the terminal, etc. The memory 20 can also be used to temporarily store data that has been output or will be output. In an embodiment, the memory 20 stores a small sample point cloud semantic segmentation program 40 based on difference enhancement, which can be executed by the processor 10, thereby implementing the small sample point cloud semantic segmentation method based on difference enhancement in the application.

[0121] The processor 10 can be a central processing unit (CPU), a microprocessor or other data processing chip in some embodiments, which is used to run program codes or process data stored in the memory 20, such as to execute the small sample point cloud semantic segmentation method based on difference enhancement, etc.

[0122] The display 30 can be, in some embodiments, an LED display, a liquid crystal display, a touch liquid crystal display, an OLED (Organic Light-Emitting Diode) touch, etc. The display 30 is used to display information at the terminal and to display a visualized user interface. The components 10-30 of the terminal communicate with each other through a system bus.

[0123] In an embodiment, when the processor 10 executes the small sample point cloud semantic segmentation program 40 based on difference enhancement in the memory 20, the steps of the small sample point cloud semantic segmentation method based on difference enhancement as described above are implemented.

[0124] The present application also provides a computer readable storage medium, wherein the computer readable storage medium stores a small sample point cloud semantic segmentation program based on difference enhancement, and when the small sample point cloud semantic segmentation program based on difference enhancement is executed by a processor, the steps of the small sample point cloud semantic segmentation method based on difference enhancement as described above are implemented.

[0125] In summary, the present application provides a small sample point cloud semantic segmentation method based on difference enhancement and related equipment, the method comprising: obtaining query set data and support set data, and performing feature alignment processing on the query set data and the support set data to obtain query set features and support set features; obtaining initial prototypes of each class according to the support set features and corresponding support set masks, wherein the initial prototypes include prototypes of multiple classes; calculating the similarity between the initial prototypes and the query set features to obtain the probability distribution of the query set features belonging to each class of foreground and background; taking the inverse of the foreground distribution according to foreground and background difference enhancement, combining it with the current background distribution to obtain an accurate background probability distribution, and integrating to obtain reliable probability distribution of the entire class; transforming the query set features through a feature aggregation module to obtain aggregated query set features; obtaining pseudo prototypes through aggregation processing of the prototype correction module according to the reliable probability distribution and the aggregated query set features; expanding the support set data according to the pseudo prototypes through the prototype correction module to obtain corrected prototypes, and obtaining the segmentation result of the point cloud according to the corrected prototypes. The present application overcomes background ambiguity through progressive foreground and background enhancement, removes outliers through a feature aggregation module under geometric guidance, and corrects prototypes using a prototype correction module, thereby obtaining better segmentation results.

[0126] It should be noted that, in the present document, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0127] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing relevant hardware (such as a processor, a controller, etc.) through a computer program, and the program can be stored in a computer readable computer readable storage medium, and the program can include the processes of the above-mentioned method embodiments when executed. The computer readable storage medium can be a memory, a magnetic disc, an optical disc, etc.

[0128] It should be understood that the application is not limited to the above examples, and those skilled in the art can make improvements or changes according to the above description, and all these improvements and changes shall fall within the protection scope of the appended claims of the present application.

Claims

1. A few-sample point cloud semantic segmentation method based on differential enhancement, characterized in that, The differential-enhanced few-sample point cloud semantic segmentation method includes: Obtain query set data and support set data, and perform feature alignment processing based on the query set data and support set data to obtain query set features and support set features; Initial prototypes for each class are obtained based on the support set features and the corresponding support set mask. The initial prototypes contain prototypes for multiple classes. Calculate the similarity between the initial prototype and the query set features to obtain the probability distribution of the query set features belonging to the foreground and background categories; The foreground distribution is inverted based on the foreground-background difference enhancement, and combined with the current background distribution to obtain an accurate background probability distribution. This is then integrated to obtain a reliable probability distribution for the entire category. The query set features are transformed by the feature aggregation module to obtain aggregated query set features. Based on the reliable probability distribution and the aggregated query set features, a pseudo-prototype is obtained through the aggregation process of the prototype correction module. The prototype correction module expands the support set data based on the pseudo-prototype to obtain the corrected prototype, and the point cloud segmentation result is obtained based on the corrected prototype.

2. The few-sample point cloud semantic segmentation method based on difference enhancement according to claim 1, characterized in that, The step of performing feature alignment processing based on the query set data and support set data to obtain query set features and support set features specifically includes: Calculate the distance between pairwise features in the query set and the support set; Invert the distance and normalize it to obtain the corresponding weight; The query set features are obtained based on the query set data, and the aligned support set features are obtained by multiplying the obtained weights with the query set features.

3. The few-sample point cloud semantic segmentation method based on difference enhancement according to claim 2, characterized in that, The process of obtaining initial prototypes for each class based on the support set features and the corresponding support set mask specifically includes: Using aligned support set features F s The initial prototypes P for each class are obtained by combining the corresponding support set mask M: The initial prototype P contains prototypes of multiple categories.

4. The few-sample point cloud semantic segmentation method based on difference enhancement according to claim 3, characterized in that, The calculation of the similarity between the initial prototype and the query set features, to obtain the probability distribution of the query set features belonging to foreground and background categories, specifically includes: Calculate the initial prototype P and the query set features F q The similarity is used to obtain the probability distribution of the query set features belonging to a specific category; For the prototype p of class i i The probability that a feature of the query set belongs to class i is defined as: in, This represents the distance between the prototype and the query set features; channel represents the number of feature channels.

5. The few-sample point cloud semantic segmentation method based on difference enhancement according to claim 4, characterized in that, The process of inverting the foreground distribution based on foreground-background difference enhancement, combining it with the current background distribution to obtain an accurate background probability distribution, and then integrating it to obtain a reliable probability distribution for the entire category, specifically includes: Based on foreground-background difference enhancement, the probability distribution of query set features belonging to different foreground and background categories is obtained. Then, the foreground distribution is inverted using foreground-background difference enhancement and combined with the current background distribution to obtain the accurate background probability distribution. Among them, Prob 0 It is the background probability distribution, and the foreground probability distribution. i Let be the probability distribution of the i-th foreground class, and C be the number of foreground classes; By combining the reliable background probability distribution and the foreground probability distribution, a reliable probability distribution for the entire category is obtained:

6. The few-sample point cloud semantic segmentation method based on difference enhancement according to claim 5, characterized in that, The process of transforming the query set features through a feature aggregation module to obtain aggregated query set features specifically includes: The geometry-guided feature aggregation module uses geometric information to select points within the k nearest neighbor range of each geometric point, for a point x within X. i Points within the k-NN range Represented as: Where q is related to x i The point closest to the k-th point; The features obtained by the feature extractor have a one-to-one correspondence with geometric points. The index of each geometric point is the same as the index of the corresponding feature. Based on the one-to-one correspondence between geometric points and features, the features corresponding to the nearest neighbors are obtained from the features. The geometry-guided feature aggregation module obtains the weights of each point relative to the initial point based on the distance between points within the k-NN and the initial point: in, Indicates the distance between two points; Based on the weights, the geometry-guided feature aggregation module aggregates local features and adds point feature information while retaining the information of the points themselves, resulting in aggregated features. For a given point feature f... i The corresponding features after aggregation Defined as: Where w represents a learnable parameter; The feature aggregation module guided by geometry will aggregate the query set features F q Transformed into aggregated query set features 7. The few-sample point cloud semantic segmentation method based on difference enhancement according to claim 6, characterized in that, The process of obtaining a pseudo-prototype by the aggregation of the reliable probability distribution and the aggregated query set features through the prototype correction module specifically includes: Based on the prototype correction module, and according to the reliable probability distribution of all categories, Prob reli Features of the query set after geometric aggregation Select the n features f with the highest probability from various probability distributions. top For category i, the selected features Defined as: Among them, prob i,j This refers to the j-th probability in the i-th class, where m is the n-th largest probability in the i-th class. Based on the one-to-one correspondence between features and probabilities, where the index of each feature is the same as the index of its corresponding probability, feature f is obtained based on these indices. top The corresponding probability prob top ; A pseudo-prototype is obtained by aggregating multiple features and their corresponding probabilities based on the prototype correction module.

8. The few-sample point cloud semantic segmentation method based on difference enhancement according to claim 7, characterized in that, The prototype correction module expands the support set data based on the pseudo-prototype to obtain a corrected prototype, and the point cloud segmentation result is obtained based on the corrected prototype, specifically including: The prototype correction module expands the support set data using pseudo-prototypes to obtain the corrected prototype: Where N is the number of data points in the support set data corresponding to the prototype category; After the corrected prototype is processed by distribution normalization and similarity calculation, the point cloud segmentation result is obtained.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a differential-enhanced few-sample point cloud semantic segmentation program stored in the memory and executable on the processor. When the differential-enhanced few-sample point cloud semantic segmentation program is executed by the processor, it implements the steps of the differential-enhanced few-sample point cloud semantic segmentation method 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 differential-enhanced few-shot point cloud semantic segmentation program, which, when executed by a processor, implements the steps of the differential-enhanced few-shot point cloud semantic segmentation method as described in any one of claims 1-8.

Images