Few-shot point cloud semantic segmentation method based on difference enhancement, and related device

Abstract

Disclosed in the present invention are a few-shot point cloud semantic segmentation method based on difference enhancement, and a related device. The method comprises: obtaining various initial prototypes on the basis of support set features and corresponding support set masks; calculating similarities between the initial prototypes and query set features, so as to obtain probability distributions of the query set features belonging to foreground and background categories; inverting foreground distributions on the basis of background-foreground difference enhancement, combining same with current background distributions to obtain accurate background probability distributions, and performing integration to obtain a reliable probability distribution of the categories as a whole; transforming the query set features by means of a feature aggregation module, so as to obtain aggregated query set features; on the basis of the reliable probability distribution and the aggregated query set features, performing aggregation by means of a prototype rectification module, so as to obtain pseudo prototypes; and on the basis of the pseudo prototypes, expanding support set data by means of the prototype rectification module, so as to obtain corrected prototypes, and obtaining a point cloud segmentation result on the basis of the corrected prototypes. The present invention achieves a better segmentation effect.

Description

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

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

[0002] Few-shot learning relies on a small amount of data and its masks. However, when dealing with multiple foreground classes, the masks of these data only indicate whether the data belongs to a specific class, and cannot indicate whether the data belongs to other foreground classes or the background. In this case, some foreground classes are misidentified as background, resulting in background ambiguity. This makes it difficult for existing solutions, including prototype networks, to distinguish between background and foreground, inevitably leading to severe prototype bias and incorrect segmentation results. Although background ambiguity has a significant impact on models, it has not received sufficient attention.

[0003] Small-sample point cloud semantic segmentation methods refer to the ability of a model to semantically segment a point cloud, given a small number of labeled new category samples, to obtain a point cloud belonging to a new category. In order to extract effective point cloud features, researchers are committed to obtaining features that can better represent category information. For example, the average value of the class feature is used as the prototype, and the point cloud is segmented by the distance between the point feature and the prototype. However, this method does not take into account the complexity of the point cloud structure itself, and the segmentation accuracy is not satisfactory. For example, by using multiple prototypes to describe complex point cloud distribution, the inference time is longer.

[0004] Existing techniques ignore the context ambiguity present in the current data. The goal of few-shot learning is to predict the label of query data given some annotated support data, which typically includes the data and the corresponding labels grouped by class. However, the foreground of one class may be mistakenly considered as background by other classes, which is called context ambiguity.

[0005] Therefore, the existing technology still needs to be improved and developed.

[0006] Summary of the Invention

[0007] The main purpose of the present invention is to provide a small sample point cloud semantic segmentation method based on difference enhancement, a terminal and a computer-readable storage medium, aiming to solve the problem that the small sample point cloud semantic segmentation method in the existing technology is prone to background ambiguity, resulting in low accuracy of small sample point cloud semantic segmentation.

[0008] To achieve the above object, the present invention provides a small sample point cloud semantic segmentation method based on difference enhancement, which comprises the following steps:

[0009] Acquire query set data and support set data, and perform feature alignment processing on the query set data and support set data to obtain query set features and support set features;

[0010] Obtaining initial prototypes of various categories according to the support set features and the corresponding support set masks, wherein the initial prototypes include prototypes of multiple categories;

[0011] Calculating 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;

[0012] According to the foreground-background difference enhancement, the foreground distribution is inverted and combined with the current background distribution to obtain an accurate background probability distribution, and then integrated to obtain a reliable probability distribution for the entire category;

[0013] Transforming the query set features through a feature aggregation module to obtain aggregated query set features;

[0014] Obtain a pseudo prototype based on the aggregation processing of the reliable probability distribution and the aggregated query set features by the prototype correction module;

[0015] The prototype correction module expands the support set data according to the pseudo prototype to obtain a corrected prototype, and obtains a point cloud segmentation result according to the corrected prototype.

[0016] Optionally, the small sample point cloud semantic segmentation method based on difference enhancement, wherein the feature alignment processing based on the query set data and the support set data to obtain the query set features and the support set features specifically includes:

[0017] Calculate the distance between each pair of features in the query set data and the support set data;

[0018] Invert the distance and normalize it to get the corresponding weight;

[0019] A query set feature is obtained according to the query set data, and the aligned support set feature is obtained by multiplying the obtained weight with the query set feature.

[0020] Optionally, in the small sample point cloud semantic segmentation method based on difference enhancement, obtaining the initial prototypes of each category according to the support set features and the corresponding support set masks specifically includes:

[0021] Using the aligned support set feature F s And the corresponding support set mask M to obtain the initial prototype P of each class:

[0022] The initial prototype P includes prototypes of multiple categories.

[0023] Optionally, in the small sample point cloud semantic segmentation method based on difference enhancement, calculating 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 specifically includes:

[0024] Calculate the initial prototype P and query set features F q The similarity of ,gets the probability distribution of query set features belonging to a specific category;

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

[0026] in, Indicates the distance between the prototype and the query set feature; channel indicates the number of feature channels.

[0027] Optionally, the small sample point cloud semantic segmentation method based on difference enhancement, wherein the foreground distribution is inverted according to the foreground-background difference enhancement, combined with the current background distribution to obtain an accurate background probability distribution, and integrated to obtain a reliable probability distribution of the entire category, specifically includes:

[0028] Based on the foreground-background difference enhancement, the probability distribution of the query set features belonging to the foreground and background categories is obtained. Based on the foreground-background difference enhancement, the foreground distribution is inverted and combined with the current background distribution to obtain the accurate background probability distribution.

[0029] Among them, Prob 0 is the background probability distribution, the foreground probability distribution Prob i is the probability distribution of the i-th foreground class, C is the number of foreground classes;

[0030] Combining the reliable background probability distribution and foreground probability distribution, we can get the reliable probability distribution of the entire category:

[0031] Optionally, in the small sample point cloud semantic segmentation method based on difference enhancement, the step of transforming the query set features through a feature aggregation module to obtain aggregated query set features specifically includes:

[0032] The feature aggregation module based on geometry guidance uses geometric information to select points within the k nearest neighbors of each geometric point, and for a point x in X i , points within the k-NN range Expressed as:

[0033] Among them, q is the same as x iThe point closest to the kth point;

[0034] The features obtained by the feature extractor have a one-to-one correspondence with the geometric points. The serial number of each geometric point is the same as the serial number of the corresponding feature. Based on the one-to-one correspondence between the geometric points and the features, the features corresponding to the neighboring points are obtained from the features.

[0035] The feature aggregation module guided by geometry obtains the weight of each point relative to the initial point based on the distance between the k-NN inner point and the initial point:

[0036] in, Indicates the distance between two points;

[0037] According to the weight, the feature aggregation module guided by geometry aggregates local features and adds 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 features after aggregation Defined as:

[0038] Among them, w represents a learnable parameter;

[0039] The query set features F are aggregated by the geometry-guided feature aggregation module q Transformed into aggregated query set features

[0040] Optionally, the small sample point cloud semantic segmentation method based on difference enhancement, wherein the pseudo prototype is obtained by aggregating the reliable probability distribution and the aggregated query set features through a prototype correction module, specifically comprising:

[0041] Based on the prototype correction module, the probability distribution Prob of all categories is reliable. reli , from the query set features after geometric aggregation Select the n features f with the highest probability in each probability distribution top , for category i, the selected features Defined as:

[0042] Among them, prob i,j It refers to the jth probability in the i-th category, and m is the nth largest probability in the i-th category. Based on the one-to-one correspondence between features and probabilities, the serial number of each feature is the same as the serial number of the corresponding probability, and the feature f is obtained according to the serial number. top The corresponding probability prob top ;

[0043] Based on the prototype correction module, multiple features and corresponding probabilities are aggregated to obtain the pseudo prototype:

[0044] Optionally, the small sample point cloud semantic segmentation method based on difference enhancement, wherein the prototype correction module expands the support set data according to the pseudo prototype to obtain a corrected prototype, and obtains a point cloud segmentation result according to the corrected prototype, specifically includes:

[0045] Based on the prototype correction module, the support set data is expanded through the pseudo prototype to obtain the corrected prototype:

[0046] Among them, N is the number of data corresponding to the prototype category in the support set data;

[0047] After the corrected prototype is subjected to distribution normalization and similarity calculation, the segmentation result of the point cloud is obtained.

[0048] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a small sample point cloud semantic segmentation program based on difference enhancement stored on the memory and runnable on the processor, and when the small sample point cloud semantic segmentation program based on difference enhancement is executed by the processor, the steps of the small sample point cloud semantic segmentation method based on difference enhancement as described above are implemented.

[0049] In addition, to achieve the above-mentioned purpose, the present invention 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.

[0050] In the present invention, query set data and support set data are obtained, and feature alignment processing is performed on the query set data and support set data to obtain query set features and support set features; initial prototypes of each category are obtained based on the support set features and the corresponding support set masks, and the initial prototypes include prototypes of multiple categories; 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 foreground and background category; the foreground distribution is inverted based on foreground and background difference enhancement, combined with the current background distribution to obtain an accurate background probability distribution, and integrated to obtain a reliable probability distribution of the entire category; the query set features are transformed 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 a pseudo prototype; the support set data is expanded by the prototype correction module based on the pseudo prototype to obtain a corrected prototype, and a point cloud segmentation result is obtained based on the corrected prototype. The present invention overcomes background ambiguity through progressive foreground and background enhancement, removes outliers through a feature aggregation module under geometry guidance, and corrects the prototype using a prototype correction module, thereby obtaining a better segmentation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] FIG1 is a flow chart of a preferred embodiment of a small sample point cloud semantic segmentation method based on difference enhancement according to the present invention;

[0052] FIG2 is a schematic diagram showing the principle of the entire semantic style process in a preferred embodiment of the semantic segmentation method of a small sample point cloud based on difference enhancement of the present invention;

[0053] FIG3 is a schematic diagram of an operating environment of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clear and distinct, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0055] To solve the existing background ambiguity problem, the present invention first uses the background-foreground difference enhancement (BFDE) strategy to use a reliable foreground to eliminate the interference information of the background, thereby obtaining a reliable foreground and background probability distribution. Then, reliable features are extracted based on the reliable probability distribution and aggregated into a pseudo-prototype. The pseudo-prototype is used to expand the features to achieve the correction of the prototype.

[0056] The small sample point cloud semantic segmentation method based on difference enhancement according to a preferred embodiment of the present invention is shown in FIG1 and FIG2 . The small sample point cloud semantic segmentation method based on difference enhancement comprises the following steps:

[0057] Step S10: Acquire query set data and support set data, and perform feature alignment processing on the query set data and support set data to obtain query set features and support set features.

[0058] Specifically, as shown in Figure 2, the feature extractor uses a dynamic graph convolutional neural network to extract features, and then performs feature alignment. Feature alignment is performed by reducing the distance between the support set and query set features. The specific steps include:

[0059] (1) Calculate the distance between each pair of features in the query set data and the support set data;

[0060] (2) Invert the distance and normalize it to obtain the corresponding weight;

[0061] (3) Use the obtained weights to multiply the query set features to obtain the aligned support set features.

[0062] Step S20: obtaining initial prototypes of various categories according to the support set features and the corresponding support set masks, wherein the initial prototypes include prototypes of multiple categories.

[0063] Specifically, as shown in Figure 2, foreground-background difference enhancement first uses the aligned support set features F s And the corresponding support set mask M to obtain the initial prototype P of each class:

[0064] The initial prototype P includes prototypes of multiple categories.

[0065] Step S30: Calculate the similarity between the initial prototype and the query set feature to obtain the probability distribution of the query set feature belonging to the foreground and background categories.

[0066] Specifically, as shown in Figure 2, the foreground-background difference enhancement calculation is performed on the initial prototype P and the query set feature F q , and thus obtain the probability distribution of the query set features belonging to a specific category.

[0067] For the prototype p of the i-th class i , the probability that the query set feature belongs to class i is defined as:

[0068] in, Indicates the distance between the prototype and the query set feature; channel indicates the number of feature channels.

[0069] Step S40: invert the foreground distribution according to the foreground-background difference enhancement, combine it with the current background distribution to obtain an accurate background probability distribution, and integrate to obtain a reliable probability distribution of the entire category.

[0070] Specifically, as shown in Figure 2, through the above operations, the foreground-background difference enhancement obtains the probability distribution of the query set features belonging to the foreground-background categories. Due to the background ambiguity problem, the current probability distribution of the background category is inaccurate. Considering that the foreground distribution is accurate, the foreground distribution is inverted based on the foreground-background difference enhancement and combined with the current background distribution to obtain the accurate background probability distribution.

[0071] Among them, Prob 0 is the background probability distribution, the foreground probability distribution Prob i is the probability distribution of the i-th foreground class, and C is the number of foreground classes.

[0072] Combining the reliable background probability distribution and foreground probability distribution, we can get the reliable probability distribution of the entire category:

[0073] Step S50: transform the query set features through a feature aggregation module to obtain aggregated query set features.

[0074] 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. i , the points within its k-NN range Expressed as:

[0075] Among them, q is the same as x i The point closest to the kth point.

[0076] Note that the features obtained by 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. Based on the one-to-one correspondence between geometric points and features, the features corresponding to the neighboring points can be obtained from the features.

[0077] Then, based on the distance between the k-NN inner point and the initial point, the feature aggregation module under geometry guidance obtains the weight of each point relative to the initial point:

[0078] in, Represents the distance between two points.

[0079] Based on the weights, the geometry-guided feature aggregation module aggregates local features and adds point feature information to retain the point information itself, thereby obtaining the aggregated features. i , the corresponding features after aggregation Defined as:

[0080] Among them, w represents a learnable parameter.

[0081] The query set features F are aggregated by the geometry-guided feature aggregation module q Transformed into aggregated query set features

[0082] Step S60: Obtain a pseudo prototype by performing aggregation processing in a prototype correction module based on the reliable probability distribution and the aggregated query set features.

[0083] Specifically, as shown in FIG2 , based on the Prototype Refinement Module (PRM), the probability distribution Prob of all reliable categories is reli , from the query set features after geometric aggregation Select the n features f with the highest probability in each probability distribution top For category i, the selected features Defined as:

[0084] Among them, prob i,j Refers to the jth probability in the i-th category, and m is the nth largest probability in the i-th category. Based on the one-to-one correspondence between features and probabilities, the serial number of each feature is the same as the serial number of the corresponding probability, and the feature f can be obtained based on the serial number. top The corresponding probability prob top .

[0085] Then, based on the prototype correction module, multiple features and corresponding probabilities are aggregated to obtain the pseudo prototype:

[0086] Step S70: The prototype correction module expands the support set data according to the pseudo prototype to obtain a corrected prototype, and obtains a point cloud segmentation result according to the corrected prototype.

[0087] Specifically, as shown in Figure 2, the prototype correction module uses pseudo-prototypes to expand the support set data to obtain the corrected prototype:

[0088] Where N is the number of data corresponding to the prototype category in the support set data.

[0089] After the corrected prototype is subjected to distribution normalization and similarity calculation, the segmentation result of the point cloud is obtained.

[0090] Among them, the present invention adjusts the feature distribution through a self-attention module to normalize the distribution:

[0091] (1) Concatenate the aggregated query set features with the corrected prototype to obtain the amplified features;

[0092] (2) Use a self-attention module to normalize the distribution of the amplified features;

[0093] (3) Split the adjusted amplified features to obtain the corresponding distribution normalization features and distribution normalization prototypes.

[0094] The present invention obtains the similarity by calculating the distance between the feature and the prototype. The similarity calculation is:

[0095] (1) Calculate the distance between the prototype and the feature;

[0096] (2) The distance is inverted and normalized to obtain the similarity between the feature and each prototype. Note that during testing, the category of the prototype with the most similar feature is selected as the prediction result, thus obtaining the segmentation result of the point cloud.

[0097] In addition, after the similarity calculation, cross-entropy loss processing will be performed according to the query set label. The similarity between the feature and the prototype is regarded as the probability that the feature belongs to these classes, and the corresponding cross-entropy loss is calculated based on the query set label and the cross-entropy function as the error of the current result.

[0098] In summary, the present invention overcomes background ambiguity through progressive foreground-background enhancement, removes outliers through a feature aggregation module guided by geometry, and corrects the prototype using a prototype correction module, thereby achieving a better segmentation effect.

[0099] After multiple groups of experimental comparisons with other small-sample point cloud semantic segmentation methods, the present invention has improved the accuracy of semantic segmentation on several commonly used datasets.

[0100] Furthermore, as shown in FIG3 , based on the above-described difference-enhanced small-sample point cloud semantic segmentation method and system, the present invention also provides a terminal, which includes a processor 10, a memory 20, and a display 30. FIG3 only shows some components of the terminal, but it should be understood that implementation of all shown components is not required, and more or fewer components may be implemented instead.

[0101] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 20 may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Furthermore, the memory 20 may also include both an internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code of the installation terminal. The memory 20 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a small sample point cloud semantic segmentation program 40 based on difference enhancement is stored on the memory 20, and the small sample point cloud semantic segmentation program 40 based on difference enhancement can be executed by the processor 10, thereby realizing the small sample point cloud semantic segmentation method based on difference enhancement in the present application.

[0102] In some embodiments, the processor 10 can be a central processing unit (CPU), a microprocessor or other data processing chip, used to run the program code or process data stored in the memory 20, such as executing the small sample point cloud semantic segmentation method based on difference enhancement.

[0103] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0104] In one 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 are implemented.

[0105] The present invention 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.

[0106] In summary, the present invention 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, performing feature alignment processing on the query set data and support set data to obtain query set features and support set features; obtaining initial prototypes of each category based on the support set features and the corresponding support set masks, the initial prototypes including prototypes of multiple categories; calculating the similarity between the initial prototypes and the query set features to obtain the probability distribution of the query set features belonging to the foreground and background categories; inverting 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 a reliable probability distribution of the entire category; transforming the query set features through a feature aggregation module to obtain aggregated query set features; aggregating the reliable probability distribution and the aggregated query set features through a prototype correction module to obtain a pseudo prototype; expanding the support set data according to the pseudo prototype through the prototype correction module to obtain a corrected prototype, and obtaining a point cloud segmentation result based on the corrected prototype. The present invention overcomes background ambiguity through progressive foreground-background enhancement, removes outliers through a feature aggregation module guided by geometry, and corrects the prototype using a prototype correction module, thereby achieving better segmentation effect.

[0107] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or terminal comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or terminal comprising the element.

[0108] Of course, those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware (such as a processor, controller, etc.) through a computer program. The program can be stored in a computer-readable storage medium that can be read by a computer. When the program is executed, it can include the processes in the above-described method embodiments. The computer-readable storage medium can be a memory, a magnetic disk, an optical disk, etc.

[0109] It should be understood that the application of the present invention is not limited to the above examples. For those skilled in the art, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A small sample point cloud semantic segmentation method based on difference enhancement, characterized in that: The small sample point cloud semantic segmentation method based on difference enhancement includes: Acquire query set data and support set data, and perform feature alignment processing on the query set data and support set data to obtain query set features and support set features; Obtaining initial prototypes of various categories according to the support set features and the corresponding support set masks, wherein the initial prototypes include prototypes of multiple categories; Calculating the similarity between the initial prototype and the query set feature to obtain the probability distribution of the query set feature belonging to the foreground and background categories; According to the foreground-background difference enhancement, the foreground distribution is inverted, combined with the current background distribution to obtain an accurate background probability distribution, and integrated to obtain a reliable probability distribution for the entire category; The query set features are transformed by a feature aggregation module to obtain aggregated query set features; According to the reliable probability distribution and the aggregated query set features, a pseudo prototype is obtained through aggregation processing by a prototype correction module; The support set data is expanded according to the pseudo prototype by the prototype correction module to obtain a corrected prototype, and a point cloud segmentation result is obtained according to the corrected prototype.

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

3. The small sample point cloud semantic segmentation method based on difference enhancement according to claim 2 is characterized in that: The initial prototypes of various types are obtained according to the support set features and the corresponding support set masks, specifically including: Using the aligned support feature F s And the corresponding support set mask M obtains the initial prototype P of each class: The initial prototype P includes prototypes of multiple categories.

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

5. The small sample point cloud semantic segmentation method based on difference enhancement according to claim 4 is characterized in that: The foreground distribution is inverted according to the foreground-background difference enhancement, combined with the current background distribution to obtain an accurate background probability distribution, and integrated to obtain a reliable probability distribution of the entire category, specifically including: Based on the foreground-background difference enhancement, the probability distribution of the query set features belonging to the foreground and background categories is obtained. Based on the foreground-background difference enhancement, the foreground distribution is inverted and combined with the current background distribution to obtain an accurate background probability distribution. Among them, Prob 0 is the background probability distribution, the foreground probability distribution Prob i is the probability distribution of the i-th foreground category, and C is the number of foreground categories; Combining the reliable background probability distribution and foreground probability distribution, we get the reliable probability distribution of the entire category:

6. The small sample point cloud semantic segmentation method based on difference enhancement according to claim 5, characterized in that: The transforming process of the query set features through a feature aggregation module to obtain aggregated query set features specifically includes: The feature aggregation module guided by geometry uses geometric information to select points within the k nearest neighbors of each geometric point. i , points within the k-NN range It is expressed as: Among them, q is related to x i The point closest to the kth point; The features obtained by the feature extractor have a one-to-one correspondence with the geometric points. The serial number of each geometric point is the same as the serial number of the corresponding feature. Based on the one-to-one correspondence between the geometric points and the features, the features corresponding to the neighboring points are obtained from the features. The feature aggregation module based on geometry guidance obtains the distance between each point and the initial point according to the distance between the k-NN inner point and the initial point. The weight of the initial point: in, Indicates the distance between two points; According to the weight, the feature aggregation module guided by geometry aggregates local features and adds 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 features after aggregation Defined as: Among them, w represents a learnable parameter; The query set features F are aggregated by the geometry-guided feature aggregation module. q Transformed into aggregated query set features 7. The small sample point cloud semantic segmentation method based on difference enhancement according to claim 6 is characterized in that: The obtaining of a pseudo prototype by performing aggregation processing by a prototype correction module based on the reliable probability distribution and the aggregated query set features specifically includes: Based on the prototype correction module, the probability distribution Prob of all categories is reliable. reli , from the query set features after geometric aggregation Select the n features f with the highest probability in each probability distribution top , for category i, the selected features Defined as: Among them, prob i,j It refers to the jth probability in the i-th category, and m is the nth largest probability in the i-th category. Based on the one-to-one correspondence between features and probabilities, the serial number of each feature is the same as the serial number of the corresponding probability. The feature f is obtained based on the serial number. top The corresponding probability prob top ; Based on the prototype correction module, multiple features and corresponding probabilities are aggregated to obtain the pseudo prototype:

8. The small 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 according to the pseudo prototype to obtain a corrected prototype, and obtains a point cloud segmentation result according to the corrected prototype, specifically including: Based on the prototype correction module, the support set data is expanded through the pseudo prototype to obtain the corrected prototype: Among them, N is the number of data corresponding to the prototype category in the support set data; After the corrected prototype is normalized and the similarity is calculated, the segmentation result of the point cloud is obtained.

9. A terminal, characterized in that: The terminal includes: a memory, a processor, and a small sample point cloud semantic segmentation program based on difference enhancement stored in the memory and executable on the processor. When the small sample point cloud semantic segmentation program based on difference enhancement is executed by the processor, the steps of the small sample point cloud semantic segmentation method based on difference enhancement as described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium, characterized in that: 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 in any one of claims 1 to 8 are implemented.

Images