CN-121982700-A - Point cloud small sample classification method, device and medium for redundant elimination and discrimination feature mining

CN121982700ACN 121982700 ACN121982700 ACN 121982700ACN-121982700-A

Abstract

The invention discloses a point cloud small sample classification method, device and medium for redundancy elimination and discrimination feature mining, which comprises the steps of S1, data acquisition, point cloud sample data acquisition from public point cloud data sources and/or actual scanning equipment and storage to storage equipment, point cloud sample data reading during training or testing, point cloud normalization and point sampling, S2, small sample task construction, S3, point feature coding, S4, three-stage point feature selection, S5, self-adaptive point budget, S6, model training, S7, model testing, S8, and result output and application. The invention realizes the suppression of the pollution of point feature redundancy and real scanning background clutter/sensing noise to global representation under the condition of small sample supervision, improves the adaptability to geometric complexity, point density change, shielding and clutter interference, thereby improving the generalization and robustness of the classification of the small point cloud samples and providing a new technical path and application prospect for intelligent perception and three-dimensional scene understanding of the point cloud.

Inventors

ZHU SHOUZHENG
YANG LIU
LIU XU
ZHANG YANGYANG
YANG WENHANG
DUAN JINYI
Wang Ceyuan
LI CHUNLAI
CHEN YUWEI

Assignees

国科大杭州高等研究院

Dates

Publication Date: 20260505
Application Date: 20260407

Claims (8)

1. A three-dimensional point cloud small sample classification method for redundant elimination and discrimination feature mining is characterized by comprising the following steps: s1, acquiring point cloud sample data from a public point cloud data source and/or actual scanning equipment and storing the point cloud sample data into storage equipment; S2, constructing a support set and a query set for small sample learning based on point cloud sample data, and forming training scene tasks and testing scene tasks which are divided by class; s3, performing point-level feature coding on the input point cloud by utilizing a backbone feature extraction network to obtain a point feature set; S4, backbone point feature selection, non-redundant feature selection and distinguishing feature mining are sequentially executed, so that a compact feature subset is constructed; S5, dynamically determining point feature budget according to the selection score marginal gain brought by the newly added point features, aggregating the compact feature subsets to form global representation, calculating a category prototype, and completing the prediction category of the query sample based on measurement matching; s6, executing steps S3-S5 on the training scene task, calculating training loss according to the prediction category and the real label, updating backbone characteristic extraction network parameters and point characteristic budget strategy parameters, and iterating until convergence to obtain a training model; S7, fixing model parameters, and executing the steps S3-S5 to output a prediction category on a test scene task; And S8, outputting the predicted category to a display device and/or a storage device, and/or inputting a downstream three-dimensional point cloud recognition and detection task.
2. The method for classifying small three-dimensional point cloud samples for redundant elimination and discrimination feature mining according to claim 1, wherein said S1 further comprises a data enhancement process in a training phase, the data enhancement process comprising at least one of random rotation, random scaling, random translation, dither noise disturbance, and point missing occlusion.
3. The method for classifying small three-dimensional point cloud samples for redundant elimination and discrimination feature mining according to claim 1, wherein the step S2 of forming training scene tasks and test scene tasks divided by classes comprises sampling N classes from class sets in each training scene task or test scene task, sampling K samples for each class to form a support set, sampling Q samples to form a query set, wherein N is an integer greater than or equal to 2, K is an integer greater than or equal to 1, and Q is an integer greater than or equal to 1.
4. The three-dimensional point cloud small sample classification method for redundant elimination and discrimination feature mining according to claim 1, wherein the step S4 comprises: S4.1, backbone point feature selection, namely selecting a preset number of point features from a point feature set according to the furthest point sampling as a backbone point feature subset, wherein the backbone point feature subset is used for representing the main geometric structure of a target; S4.2, selecting non-redundant point features, namely judging the non-backbone point features as redundant and eliminating the non-backbone point features when the similarity between the non-backbone point features and any backbone point feature in the backbone point feature subsets is larger than a preset threshold value, so as to obtain the non-redundant point feature subsets; And S4.3, mining the distinguishing features, namely calculating the similarity between the non-redundant point feature subset and the support set category prototype to obtain the distinguishing score of the non-redundant features, screening the non-redundant point features according to the distinguishing score, and reserving the point features with higher distinguishing scores to form a distinguishing point feature subset.
5. The three-dimensional point cloud small sample classification method for redundant elimination and discrimination feature mining according to claim 1, wherein the adaptive point budget in S5 comprises: s5.1, gradually expanding the point features, namely gradually adding the point features into the compact feature subset from high to low according to the point feature selection scores; s5.2, calculating a marginal gain, namely calculating a selected score increment of the compact feature subset after each point feature addition as the marginal gain; and S5.3, self-adaptive stopping, namely stopping adding the point characteristic and determining the point characteristic budget of the sample when the marginal gain of the point characteristic is continuously added for a plurality of times to meet the preset stopping condition.
6. The method for classifying small three-dimensional point cloud samples for redundancy elimination and discrimination feature mining according to claim 1, wherein the class prototypes are mean prototypes for supporting global representation of the same class of samples in a centralized manner, the metric matching comprises the steps of calculating monotone transformation of cosine similarity, inner product similarity or Euclidean distance between the global representation of the query sample and each class of prototypes, and selecting a class with the largest similarity as a prediction class.
7. The three-dimensional point cloud small sample classification device for redundancy elimination and discrimination feature mining is characterized by comprising one or more processors, wherein the three-dimensional point cloud small sample classification device is used for realizing the three-dimensional point cloud small sample classification method according to any one of claims 1-6.
8. A computer-readable storage medium, having stored thereon a program which, when executed by a processor, implements the three-dimensional point cloud small sample classification method of any of claims 1-6.

Description

Point cloud small sample classification method, device and medium for redundant elimination and discrimination feature mining Technical Field The invention relates to the technical field of three-dimensional point cloud small sample learning and three-dimensional computer vision, in particular to a point cloud small sample classification method, device and medium for redundant elimination and discrimination feature mining. Background In the field of three-dimensional computer vision, point cloud classification is an important basis for three-dimensional scene understanding, and is widely applied to scenes such as automatic driving, robot navigation, three-dimensional perception and the like. Because the point cloud has disorder and unstructured characteristics, and the real scanning point cloud is often accompanied by uneven point density, shielding loss, background clutter and sensing noise interference, the robustness and generalization of the point cloud identification model in a complex environment are limited. Meanwhile, the high-quality and task-related point cloud data with labels in the real scene are high in acquisition cost, new classes continuously appear but only a small number of label samples can be provided, so that the classification of small point cloud samples becomes a key requirement. The existing small sample point cloud classification method generally adopts a model of feature encoder and metric-based prototype matching, namely, extracting point cloud features by a backbone network, and then carrying out similarity matching on query sample features and each class of prototypes to finish classification. To enhance representation capabilities, the prior art generally tends to preserve and exploit more point-level features as much as possible, or to reduce information loss by multi-view projection/multi-modal fusion/improvement aggregation modules. However, in real scan data, the point cloud often contains a large number of redundant points, background clutter and outliers, and under the condition of small sample supervision, the redundancy and noise are easier to introduce interference in the feature aggregation process, dilute discriminant clues and pollute global representation, thereby influencing the generalization performance of the model. Therefore, a method is needed that can effectively reject redundant point features and suppress noise interference under a small sample point cloud classification framework, simultaneously excavate key point features related to category discrimination, and can adaptively determine point feature budget according to geometric complexity, density change and shielding degree of different samples. Disclosure of Invention The invention aims to provide a point cloud small sample classification method, device and medium for redundancy elimination and discrimination feature mining, which are used for solving the problems that in the prior art, point cloud classification tasks of small samples are high in point feature redundancy, aggregation representation is polluted due to real scanning background clutter and noise, and fixed point budgets are difficult to adapt to different geometric complexity and density changes. In order to achieve the above purpose, the present invention provides the following technical solutions: A three-dimensional point cloud small sample classification method for redundant elimination and discrimination feature mining comprises the following steps: s1, acquiring point cloud sample data from a public point cloud data source and/or actual scanning equipment and storing the point cloud sample data into storage equipment; S2, constructing a support set and a query set for small sample learning based on point cloud sample data, and forming training scene tasks and testing scene tasks which are divided by class; s3, performing point-level feature coding on the input point cloud by utilizing a backbone feature extraction network to obtain a point feature set; S4, backbone point feature selection, non-redundant feature selection and distinguishing feature mining are sequentially executed, so that a compact feature subset is constructed; S5, dynamically determining point feature budget according to the selection score marginal gain brought by the newly added point features, aggregating the compact feature subsets to form global representation, calculating a category prototype, and completing the prediction category of the query sample based on measurement matching; s6, executing steps S3-S5 on the training scene task, calculating training loss according to the prediction category and the real label, updating backbone characteristic extraction network parameters and point characteristic budget strategy parameters, and iterating until convergence to obtain a training model; S7, fixing model parameters, and executing the steps S3-S5 to output a prediction category on a test scene task; And S8, outputting the predicted category to a display