CN-121982424-A - Uncertainty guided analytic type high-resolution range profile category increment learning method

CN121982424ACN 121982424 ACN121982424 ACN 121982424ACN-121982424-A

Abstract

The application provides an uncertainty-guided analytic type high-resolution range profile class increment learning method, which mainly solves the problem that catastrophic forgetting is easy to occur in radar high-resolution range profile object recognition and old class recognition performance is obviously degraded. The method comprises the steps of constructing a multi-resolution structured feature representation, using a statistical prototype comprising mean and second order statistics as a unique carrier of old class knowledge, describing old class distribution characteristics under the condition of no original sample by recursively updating variance or covariance information, carrying out progressive analytic type updating on a classifier based on a recursion least square closed solution of covariance matrix recursion tracking to replace counter propagation training, introducing a self-adaptive prototype distillation and uncertainty guiding mechanism based on the statistical prototype, and realizing old class knowledge maintenance in the incremental training process. Experimental results show that the method can keep high and stable overall recognition accuracy in the target category continuous expansion scene under the condition that the old category sample is not used at all.

Inventors

WANG PENGHUI
LIU HONGWEI
CHANG ZHIPENG
GUO SHUAI
CHEN BO
CHEN WENCHAO
LI YUXIN

Assignees

西安电子科技大学

Dates

Publication Date: 20260505
Application Date: 20260210

Claims (10)

1. An uncertainty-guided analytic type high-resolution range profile category increment learning method is characterized by comprising the following steps of: Step 1, in an initial training stage, training a one-dimensional feature extraction network by using initial base class HRRP data, and freezing parameters of the one-dimensional feature extraction network after the training is finished, and initializing an analytic classifier, wherein the analytic classifier comprises a main classifier and a residual classifier; And 2, in an incremental learning stage, inputting new HRRP data of the current stage into the frozen feature extraction network, constructing a multi-resolution structured feature representation, constructing and updating a statistical prototype without an old sample, updating an analytical main classifier based on covariance tracking, performing uncertainty-guided cognitive uncertainty weighted fusion, performing self-adaptive prototype distillation and old class protection, performing structural residual decomposition and residual classifier updating, and performing competitive old class inhibition, thereby outputting the prediction classification of the new HRRP data.
2. The uncertainty-guided, resolved-high-resolution-range-profile-category-increment learning method of claim 1, wherein step 1 comprises: step 1.1, acquiring HRRP one-dimensional distance image data used in an initial training stage, formatting the HRRP one-dimensional distance image data to enable the HRRP one-dimensional distance image data to meet the input requirement of a one-dimensional convolutional neural network, and obtaining initial base class HRRP data; step 1.2, inputting the HRRP data of the initial base class into a one-dimensional ResNet < 18 > feature extraction network, and performing end-to-end training on network parameters by adopting a back propagation-based mode; And 1.3, after the one-dimensional ResNet-18 feature extraction network training is finished, freezing the internal parameters of the one-dimensional ResNet-18 feature extraction network training, so that the one-dimensional ResNet-18 feature extraction network training is not involved in back propagation updating in an incremental learning stage, and initializing an analytical classifier for a subsequent incremental learning stage.
3. The uncertainty-guided, analytic high resolution range profile category delta learning method of claim 1, wherein the multi-resolution structured feature representation construction comprises: step 2.1, extracting basic features from the new class of HRRP data by utilizing a one-dimensional feature extraction network of frozen parameters, and outputting a feature map composed of the basic features; Step 2.2, carrying out downsampling treatment on the feature map in different scales to obtain multi-scale feature representation; and 2.3, carrying out weighted combination on the multi-scale feature representation to obtain multi-resolution structured features.
4. The uncertainty-guided, high-resolution range profile class increment learning method of claim 1, wherein the statistical prototype construction and updating without old samples comprises: step 3.1, calculating the sample feature mean value of each new category in the incremental learning stage based on the multi-resolution structured feature representation, and taking the sample feature mean value as the mean value of a new category statistical prototype; step 3.2, for the learned old category, keeping the mean value of the statistical prototype unchanged; and 3.3, for the new category, calculating the characteristic variance of the current batch of samples to update the statistical prototype variance of the current batch of samples, for the old category, recursively updating the statistical prototype variance of the old category by using a combined variance formula, and performing lower-bound cutoff processing on the updated variance after updating.
5. The uncertainty-guided, analytic high resolution range profile category delta learning method of claim 1, wherein the covariance tracking-based analytic master classifier update comprises: step 4.1, projecting the multi-resolution structured feature representation through a random buffer area and inputting the multi-resolution structured feature representation into the main classifier; Step 4.2, introducing and maintaining a covariance matrix for recursively tracking statistical information of the historical input features; And 4.3, based on a recursive least square analytic solution, combining the covariance matrix with new class characteristics of the current batch and labels thereof, performing closed solving and updating on the weight matrix of the main classifier, and completing gradient-free back propagation analytic weight updating.
6. The uncertainty-guided, resolved-high-resolution range profile class-increment learning method of claim 1, wherein the uncertainty-guided, cognitive uncertainty-weighted fusion comprises: step 5.1, calculating a cognitive uncertainty value of the main classifier on the sample based on the introduced covariance matrix; Step 5.2, carrying out normalization processing on the cognitive uncertainty values of all samples in the current batch, and converting the normalized result into a fusion weight through a Sigmoid function; And 5.3, based on the fusion weight, carrying out self-adaptive fusion on the main classifier, the prototype classifier based on the statistical prototype and the residual classifier, and dynamically adjusting the fusion weights of the prototype classifier and the residual classifier according to the progress of the incremental task in the fusion process.
7. The uncertainty-guided, resolved high-resolution range profile class delta learning method of claim 1, wherein the adaptive prototype distillation and old class protection comprises: step 6.1, calculating the calibration distance from the current sample characteristic to each old class prototype according to the mean value and the variance in the old class statistical prototype, and converting the calibration distance into a soft label for representing the attribution probability of the old class; Step 6.2, dynamically calculating and restricting the numerical range of the distillation weight according to the task progress determined by the ratio of the learned old category number to the current stage newly added category number; Step 6.3, when training the main classifier, mixing the original single-heat coded label with the soft label according to the distillation weight for the old class to form a final training supervision signal; And 6.4, in the reasoning stage, carrying out weighted fusion on the original prediction log probability of the old category and the corresponding calibration distance so as to enhance the confidence coefficient of the prediction of the old category.
8. The uncertainty-guided, resolved-range-profile class-increment learning method of claim 1, wherein the structural residual decomposition and residual classifier update comprises: 7.1, carrying out structural residual decomposition on the multi-resolution structured feature representation to obtain a low-frequency structural component and a high-frequency disturbance component; And 7.2, projecting the high-frequency disturbance component through a random buffer area and inputting the high-frequency disturbance component into the residual classifier so as to judge and learn new classes.
9. The uncertainty-guided, resolved-range-profile class-increment learning method of claim 1, wherein the competitive old-class suppression comprises: Step 8.1, identifying the old class with the highest competitive power in each sample based on the prototype calibration score; step 8.2, calculating confidence margins of the new class and the old class, and inhibiting the old class classifier when the margins are positive; Step 8.3, dynamically adjusting the inhibition intensity according to the task progress; and step 8.4, suppressing the old class classifier with the highest competitiveness based on the suppression intensity.
10. A system for implementing the uncertainty-guided, resolved-high-resolution range profile class-increment learning method of any one of claims 1-9, comprising: A memory for storing a computer program; a processor for implementing the functions of the following modules when executing the computer program: The device comprises an initial training module, a first analysis type classifier and a second analysis type classifier, wherein the initial training module is used for training a one-dimensional feature extraction network by utilizing initial base class HRRP data, and freezing parameters of the one-dimensional feature extraction network after the training is finished, and initializing the analysis type classifier, wherein the analysis type classifier comprises a main classifier and a residual classifier; The incremental learning module is used for inputting new HRRP data of the current stage into the frozen feature extraction network, constructing a multi-resolution structured feature representation, constructing and updating a statistical prototype without an old sample, updating an analytical main classifier based on covariance tracking, performing uncertainty-guided cognitive uncertainty weighted fusion, performing self-adaptive prototype distillation and old class protection, performing structural residual decomposition and residual classifier updating and performing competitive old class inhibition, and outputting the prediction classification of the new HRRP data.

Description

Uncertainty guided analytic type high-resolution range profile category increment learning method Technical Field The application belongs to the technical field of radar target recognition, and particularly relates to an uncertainty-guided analytic type high-resolution range profile class increment learning method. Background The radar one-dimensional range profile (High Resolution Range Profile, HRRP) is a one-dimensional feature obtained by performing range-wise high-resolution processing on the target echo signal, and can reflect the scattering point distribution structure of the target in the radial direction. Because the HRRP has the characteristics of convenience in acquisition, insensitivity to illumination and meteorological conditions and the like, the HRRP is widely applied to the fields of radar target identification, battlefield situation awareness and the like, and is particularly suitable for realizing rapid and steady target identification in complex electromagnetic environments and dynamic task scenes. In actual combat or surveillance tasks, radar systems are often faced with a continuously increasing demand for target types, such as the need to identify new emerging aircraft, ships, etc. targets. This requires that the recognition model not be used only after training on a fixed set of categories, but rather has the ability to learn new categories continuously while not forgetting old categories, i.e., class-increment learning (Class-INCREMENTAL LEARNING, CIL) capability. However, the core scientific difficulty faced by class incremental learning is "catastrophic forgetting", in which when a model is updated with only new class samples, its parameter optimization process tends to cover or destroy old class feature representations and decision boundaries learned before, resulting in a dramatic drop in recognition performance for the old class. This problem becomes particularly acute in certain engineering scenarios with stringent constraints. For example, in applications involving sensitive data or onboard limited storage/computing platforms, such as military radars, unmanned reconnaissance, systems often do not allow, nor are it difficult to store and replay, for long periods of time, the historically acquired raw echo data, due to data security concerns or hardware resource limitations. This means that at each stage of incremental updating of the model, only the target data newly acquired at the current stage can be accessed, and no real sample of any old class of history can be obtained for review or playback. Under the strict constraint of 'no old sample playback', how to effectively inhibit catastrophic forgetting becomes a very challenging technical problem in the field of radar HRRP identification. To address this challenge, the prior art has proposed several classes of incremental learning schemes. Among them, the representative solution closest to the present application is a radar HRRP continuous learning method based on generation of a countermeasure network (GAN) (as disclosed in chinese patent CN 202111239763.8). The method has the core ideas that the self-encoder is separated by the condition and the data distribution of historical tasks is learned by a countermeasure network (CVAE-GAN) is generated, so that pseudo samples of historical categories are generated, and the pseudo samples are mixed with real new data in an incremental training stage to simulate the effect of 'data playback', so that the model is expected to review old knowledge while learning new. However, such methods based on generative playback have significant drawbacks in radar HRRP application scenarios: First, the physical consistency and reliability of the generated samples is difficult to guarantee. The radar HRRP signal deeply contains a physical scattering mechanism of a target structure and a material, and the data distribution of the radar HRRP signal is essentially different from that of a natural image. The generative model is extremely prone to produce "artifacts" that do not conform to the true scattering law in the absence of explicit physical constraints. The participation of these distorted pseudo-samples in training can lead to uncontrolled shifts in classifier decision boundaries, thereby compromising the recognition accuracy and generalization ability of the model. Second, the training process is complex and unstable. The generation of the countermeasure training has the problems of mode collapse, difficult convergence and the like, and although the self-encoder of the variation is combined to improve the stability, the training process is still difficult to accurately control and the engineering parameter adjustment difficulty is high when the HRRP is fitted with the complex, high-dimensional and possibly azimuth sensitive signal distribution. Third, the system complexity and computational overhead are high. The method needs to synchronously maintain and update a plurality of large-scale net