CN-121999277-A - Strong classification method combining boundary feature generation and hypersphere constraint

CN121999277ACN 121999277 ACN121999277 ACN 121999277ACN-121999277-A

Abstract

The invention discloses a method for strongly classifying boundary feature generation and hypersphere constraint, which comprises the steps of training a prototype network by using samples of known types in a library, learning a prototype for each category to initialize a feature space, sorting depth features extracted by a feature extraction network according to the distance between the depth features and the prototype, screening feature points close to a classification boundary, applying directional disturbance to generate difficultly-classified sample features near the decision boundary, applying hypersphere constraint to original features and the generated difficultly-classified sample features, wherein the constraint explicitly limits the similar features in a compact area in the hypersphere, ensures effective control of compactness of the features in the category and inter-category intervals from the geometric structure, provides a containing space for potential unknown similar abnormal-shaped target features, improves the generalization capability of the model to the unknown similar objects, and performs cooperative optimization on the feature extraction network and the learnable prototype by combining prototype contrast loss to finish classification of the unknown similar objects.

Inventors

CHEN JIAN
WANG JIANWEI
DU LAN

Assignees

西安电子科技大学

Dates

Publication Date: 20260508
Application Date: 20260107

Claims (10)

1. A method for classifying by combining boundary feature generation and hypersphere constraint is characterized by comprising the following steps: Acquiring radar HRRP echo data of a target to be detected; Classifying radar HRRP echo data of the target to be detected by utilizing a pre-trained generalized classification model based on boundary feature generation and hyperspherical constraint to obtain the class of the target to be detected, The training process of the generalized classification model based on boundary feature generation and hypersphere constraint comprises the following steps: Collecting radar HRRP echo data of multiple types of targets, and constructing a training set according to the major types of the targets; Inputting training samples in a training set into a prototype network constructed based on a convolutional neural network, and performing preliminary training on the prototype network by utilizing a prototype comparison loss function to obtain the prototype network with basic classification capability; Inputting all training samples in a training set into a prototype network with basic classification capability, sorting according to the distance between each training sample and the prototype of the category, and selecting low confidence sample characteristics; Combining the characteristics of the difficult-to-separate samples with the training samples corresponding to the training set to obtain a combined training set, training the prototype network with basic classification capacity again based on the combined training set, and introducing hypersphere constraint in the training to construct a strong relaxation characteristic space so as to obtain a pre-trained generalized classification model based on boundary characteristic generation and hypersphere constraint.
2. The method for the robust classification of boundary feature generation and hypersphere constraint according to claim 1, wherein in the process of performing preliminary training on the prototype network, the prototype network uses a multi-layer convolution operation to realize nonlinear transformation, training samples are mapped to a feature space with discriminant, in the feature space, the prototype network learns a representative prototype vector for each target class, and sample features of the same class are close to corresponding class prototypes, and sample features of different classes are far away from each other by optimizing a prototype comparison loss function.
3. A method of generalized classification in combination with boundary feature generation and hypersphere constraints as recited in claim 1, the prototype contrast loss function is characterized by comprising the following expression: ; Wherein, the A prototype contrast loss function is represented and, Representing the total number of categories, Represent the first The class is used for the total number of target samples for training, Represent the first Class III Deep-level features of the individual samples, Represent the first A class prototype of a class is provided, Representing the distance between the proposed deep feature vector and the class prototype.
4. The method for the robust classification of boundary feature generation and hypersphere constraint according to claim 1, wherein the step of inputting all training samples in a training set into a prototype network with basic classification capability, sorting according to the distance between each training sample and the prototype of the class, and selecting low-confidence sample features comprises the following steps: Inputting all training samples in a training set into a prototype network with basic classification capability, and acquiring deep features corresponding to each training sample through forward propagation; in the learned feature space, calculating Euclidean distance between each deep feature and a large-class prototype corresponding to each deep feature; And sequencing all training samples according to Euclidean distances corresponding to each deep feature, and selecting deep features of part of the training samples according to the sequencing order as low-confidence sample features.
5. The method of claim 1, wherein the applying a directional perturbation to the low confidence sample features to generate refractory sample features at classification boundaries comprises: Confirming a classification boundary distance corresponding to the low confidence sample feature based on a distance formula from the point to the hyperplane; And selecting the classification boundary distance with the smallest modulus as a disturbance vector, applying the disturbance vector to the low-confidence sample characteristics, and generating the difficult-to-classify sample characteristics positioned at the classification boundary.
6. The method for categorizing a combination of boundary feature generation and hypersphere constraints as described in claim 5, the method is characterized in that the expression of the classification boundary distance is as follows: ; Wherein, the Representing categories Low confidence sample features of (2) Category and category Is defined by a classification boundary distance of (a), Representing categories And Is used to determine the decision boundary of (c), Representing categories And Is determined by the normal vector of the decision boundary of (c), , Representing categories And Is used to determine the bias term of the decision boundary of (c), , Representing categories Is a model of the class of the model, Representing categories Is a category prototype of (a).
7. The generalized classification method combining boundary feature generation and hypersphere constraint according to claim 1, wherein the expression of the total loss function of the generalized classification model based on boundary feature generation and hypersphere constraint is as follows: ; Wherein, the Representing the total loss function of the device, Representing prototype losses The weight in the total loss is calculated, Representing the loss of the prototype, Representing hypersphere constraint loss The weight in the total loss is calculated, Representing the hyperspherical constraint loss.
8. The method for generalized classification of boundary feature generation and hypersphere constraint as recited in claim 7, wherein the expression of hypersphere constraint loss is as follows: ; Wherein, the Representing the loss of the constraint of the hypersphere, Representing the total number of categories, Represent the first The class is used for the total number of target samples for training, Represent the first Class III Deep-level features of the individual samples, The representation is used to measure the distance between the proposed deep feature vector and the class prototype, Represent the first A prototype representation of the class, Representing a preset learnable radius.
9. An electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory are in communication with each other through the communication bus; The memory is used for storing a computer program; The processor is configured to implement the method for generalized classification combining boundary feature generation and hypersphere constraint according to any one of claims 1 to 8 when executing the program stored on the memory.
10. A computer-readable storage medium comprising, The computer readable storage medium has stored thereon a computer program which, when executed by a processor, implements the steps of a method for enhanced classification of boundary feature generation and hypersphere constraints as claimed in any one of claims 1 to 8.

Description

Strong classification method combining boundary feature generation and hypersphere constraint Technical Field The invention belongs to the technical field of radar target classification, and particularly relates to a generalized classification method combining boundary feature generation and hypersphere constraint. Background In modern scene perception, the intelligent radar has great application value for quick perception, and judgment aiming at the target category is a key information technology of the intelligent radar. The technology can realize automatic classification, identification and tracking of the target. Compared with a narrowband radar signal, the broadband radar signal can provide more abundant target structure information, so that a finer basis is provided for target identification. In the broadband radar signal, radar High-Resolution Range Profile, HRRP has become one of the important means for target identity judgment by virtue of its easy acquisition, processing and low requirements on radar systems. However, when the complex target perception task is oriented, the target radar echo cannot be acquired in advance to build a library due to complex target models, so that similar abnormal targets similar to the training target but different in model cannot be accurately classified in practical application, accurate evaluation of the target is further affected, and decision errors are caused. Therefore, the classification judgment of the targets with the model which is not seen is realized, and the types of the targets can be effectively prejudged. The existing class judgment technology based on HRRP is mostly aimed at target individual model identification, and the essence is that a model is required to 'see' all possible models, which is not only unrealistic in practical application, but also fails to tighten the core requirement that a target large class is focused on rather than a specific model. The core of the large class classification of targets is that only the large class labels of the targets are used for training without depending on specific model labels, the fundamental difficulty is how to let a model learn to ignore the differences among different models in the class, and the generalization capability of the targets with unknown models is shown in the classification process so as to cope with the targets with new models which are never seen. The class judgment method based on the radar HRRP target can be divided into a target class judgment method based on traditional machine learning and a class judgment method based on a deep neural network. HRRP category judgment methods based on traditional machine learning typically rely on a priori information for human selection to construct features for identification and for subsequent classification tasks. The HRRP category judging method based on deep learning realizes approximation of complex functions between the input radar target echo and the category labels through multi-layer nonlinear transformation, and end-to-end learning between input data and output categories is completed. The existing radar HRRP target class judging technology focuses on target individual model identification, and researches on the major class classification problem of targets are relatively weak. While the traditional machine learning method has the built tree in the fine recognition of the target model, the traditional machine learning method relies on the prior knowledge or the artificial design characteristics of the target, when the problem of classifying the unknown target is faced, the essential characteristics with generalization capability on the unknown model are difficult to extract from the limited large-class samples, so that the limitation in the large-class classification task is obvious, while the mainstream deep learning method can automatically learn the characteristics, but the supervision signals often drive the model to pay attention to the detailed characteristics of the specific model rather than learn the large-class commonality of the similar abnormal-shaped target, so that the learned decision boundary is too compact, and enough characteristic space cannot be reserved for the similar abnormal-shaped target which is not seen during training, so that the generalization capability in the large-class classification task is obviously limited. Disclosure of Invention In order to solve the problems in the prior art, the invention provides a generalized classification method combining boundary feature generation and hypersphere constraint. The technical problems to be solved by the invention are realized by the following technical scheme: in a first aspect, the present invention provides a method for generalized classification combining boundary feature generation and hypersphere constraint, comprising: Acquiring radar HRRP echo data of a target to be detected; Classifying radar HRRP echo data of the target to be detected by utilizing a