A radar target recognition method based on self-supervised model transfer learning

Abstract

The present invention discloses a radar target recognition method based on self-supervised model transfer learning, which includes two stages. In the first stage, the backbone network in the self-supervised model is pre-trained on a large number of unlabeled image samples in a self-supervised manner, and the network weights are updated through similarity loss to maximize the similar features of the same-class targets and weaken the similar features of different-class targets, so as to effectively extract the similar features of the same-class targets by the network. In the second stage, the network parameters of the backbone network in the pre-trained self-supervised model are transferred, and a small number of labeled radar samples are used to fine-tune the network to achieve radar target recognition. The method of the present invention can effectively identify target features and recognize target categories, solve the problem of insufficient network learning caused by the scarcity of radar image sample labels, improve the recognition performance of the network for radar targets under a small number of labeled data samples, and improve the accuracy and robustness of the small-sample radar target recognition network compared with other target recognition methods.

Description

Technical Field

[0001] The present invention belongs to the technical field of target recognition, and particularly relates to a radar target recognition method based on self-supervised model transfer learning. Background Art

[0002] In remote sensing image processing, the main information detected by imaging radar microwave remote sensing is the microwave information reflected after the interaction between the target and the electromagnetic wave in the microwave band. Therefore, the radar images obtained by the imaging radar contain the shape and scattering information of the target object, and can well characterize specific target features. Among many imaging radars, synthetic aperture radar (SAR) is an important microwave remote sensing imaging system, which can collect all-weather and high-resolution images day and night. And SAR automatic target recognition (ATR) technology is one of the key challenges in SAR applications. Therefore, the application of SAR ATR in the civilian field has important significance.

[0003] The emergence of deep learning has greatly promoted the development of image processing applications. Among them, the Convolutional Neural Network (CNN) analyzes and learns visual features from a large amount of data. Thanks to its powerful feature learning ability, CNN has also made certain progress in the field of SAR ATR. However, such methods require a large amount of labeled data sets to train the model, while most SAR image data sets are unlabeled or sparsely labeled, which will cause overfitting when training deep networks. The literature "Chen, Sizhe, et al. 'Target classification using the deep convolutional networks for SAR images.' IEEE Transactions on Geoscience and Remote Sensing 54.8 (2016): 4806 - 4817." proposed a fully convolutional network that replaces all fully connected layers with convolutional layers. This method reduces overfitting by reducing model parameters, but its recognition performance decreases significantly with the reduction of labeled training samples. Since SAR images are more difficult to obtain than optical images, and manually labeling SAR image data is time-consuming and laborious, only a small part of the existing radar image data is labeled data. Therefore, in the field of SAR ATR, it is quite challenging to train deep networks using supervised learning.

[0004] Self-supervised learning learns the general features of data by estimating the similarity between targets. This type of learning can classify similar targets without prior information and distinguish them from different targets. The literature "Ciga, Ozan, Tony Xu, and Anne Louise Martel. 'Self supervised contrastive learning for digital histopathology.' Machine Learning with Applications 7 (2022): 100198." and the literature "Güldenring, Ronja, and Lazaros Nalpantidis. 'Self-supervised contrastive learning on agricultural images.' Computers and Electronics in Agriculture 191 (2021): 106510." show that self-supervised learning has achieved certain achievements in both digital medical pathology images and agricultural images. If the ability of the self-supervised model to extract similarity features can be fully utilized, self-supervised learning has broad prospects in promoting small-sample SAR image target recognition. Summary of the Invention

[0005] To solve the above technical problems, the present invention proposes a radar target recognition method based on self-supervised model transfer learning. Aiming at the problem of limited labeled data samples in SAR image target recognition, a new framework is proposed for training a deep neural network for SAR target recognition to eliminate the need for a large amount of labeled training data.

[0006] The technical solution adopted by the present invention is as follows: A radar target recognition method based on self-supervised model transfer learning, the specific steps are as follows:

[0007] S1. Pre-train a self-supervised model to improve the effectiveness of feature extraction;

[0008] In the self-supervised model, the backbone network is pre-trained on a large number of unlabeled image samples in a self-supervised manner, and the network weights are updated using the similarity loss to maximize the similar features of the same-class targets and weaken the similar features of different-class targets, so as to realize the effective extraction of the similar features of the same-class targets by the network.

[0009] Among them, the loss used for self-supervised learning in the pre-training stage is called the similarity loss.

[0010] When a pair of images is input into the self-supervised model, if they are similar, the model infers them as 1, otherwise as 0.

[0011] S2. Fine-tune the classification and recognition network to achieve radar image target recognition;

[0012] Transfer the network parameters of the backbone network in the pre-trained self-supervised model, add a classifier behind it, input the labeled radar target image sample x, and assume there are N c types of radar targets, then use an N c -dimensional one-hot vector y to represent the class label. The output of this classification and recognition network is the conditional probability distribution of the class label, which can be expressed by the following formula:

[0013]

[0014] where, θ represents a series of updatable parameters including weights w and biases b; the added classifier includes operations of average pooling, Flatten, fully connected layer, and SoftMax.

[0015] Use cross-entropy loss to fine-tune the classification and recognition network, thereby achieving radar image target recognition. The loss function can be expressed by the following formula:

[0016]

[0017] where, y c represents the true one-hot label, and f c (x,θ) represents the feature mapping function of the classifier.

[0018] Advantages of the present invention: The method of the present invention includes two stages. In the first stage, the backbone network in the self-supervised model is pre-trained on a large number of unlabeled image samples in a self-supervised manner. The network weights are updated through similarity loss, maximizing the similar features of the same-class targets and weakening the similar features of different-class targets, so as to effectively extract the similar features of the same-class targets by the network. In the second stage, the network parameters of the backbone network in the pre-trained self-supervised model are transferred, and the network is fine-tuned using a small number of labeled radar samples to achieve radar target recognition. The method of the present invention can effectively identify target features and recognize target categories, solve the problem of insufficient network learning caused by the scarcity of radar image sample labels, improve the recognition performance of the network for radar targets under a small number of labeled data samples, and compared with other target recognition methods, improve the accuracy and robustness of the small-sample radar target recognition network. Description of the Drawings

[0019] Figure 1 It is a flowchart of a radar target recognition method based on self-supervised model transfer learning of the present invention.

[0020] Figure 2This is a comparison chart of the final recognition performance obtained by simulating the method of the present invention using five classic self-supervised models in the embodiments of the present invention. Detailed implementation manners

[0021] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0022] As Figure 1 shown, a flowchart of a radar target recognition method based on self-supervised model transfer learning of the present invention is as follows:

[0023] S1. Pre-train a self-supervised model to improve the effectiveness of feature extraction;

[0024] In the self-supervised model, the backbone network is pre-trained on a large number of unlabeled image samples in a self-supervised manner, and the network weights are updated using the similarity loss to maximize the similar features of the same class of targets and weaken the similar features of different classes of targets, so as to realize the effective extraction of the similar features of the same class of targets by the network.

[0025] Among them, the loss used for self-supervised learning in the pre-training stage is called the similarity loss, and its purpose is to distinguish the features of the input vectors.

[0026] When a pair of images is input into the self-supervised model, if they are similar, the model infers them as 1, otherwise as 0.

[0027] In the simulation experiment of this embodiment, five classic self-supervised models are adopted, namely BYOL, DenseCL, MoCoV2, SimCLR, and SimSiam. Table 1 lists the learning rate settings of different self-supervised models in the pre-training stage.

[0028] Table 1

[0029]

[0030] Through the pre-training process of step S1, the network model can effectively extract target features, maximize the similar features between the same class of targets, and weaken the similar features between different classes of targets. Then, through the network training fine-tuning in step S2, the intra-class features of the targets are aggregated and the inter-class features are separated.

[0031] As Figure 1As shown in the figure, the framework of the method of the present invention includes two training stages. The first stage is the pre-training stage, where the backbone network is trained on the large-scale unlabeled natural image dataset ImageNet in a self-supervised manner. By using the similarity loss, while maximizing the similar features of the same-class targets, weakening the similar features of different-class targets, and improving the effectiveness of the extracted features. The second stage is the fine-tuning stage, where the parameter weights of the backbone network in the self-supervised network model are transferred to the target classification network, and the classification network is fine-tuned by inputting a small number of labeled radar data samples, and finally the accurate recognition of small-sample SAR targets under the transfer learning of the self-supervised model is realized.

[0032] S2. Fine-tune the classification and recognition network to achieve radar image target recognition;

[0033] Transfer the network parameters of the backbone network in the pre-trained self-supervised model, add a classifier behind it, input the labeled radar target image sample x, and assume there are N c classes of radar targets, then use an N c -dimensional one-hot vector y to represent the class label. The output of this classification and recognition network is the conditional probability distribution of the class label, which can be expressed by the following formula:

[0034]

[0035] where, θ represents a series of updatable parameters including weights w and biases b; the added classifier includes operations of average pooling, Flatten, fully connected layer, and SoftMax. The average pooling operation refers to taking the average value of the feature points in the neighborhood of the feature map to reduce the error of the increased variance of the estimated value caused by the limited neighborhood size; the Flatten operation pulls the input multi-dimensional data into one dimension, which is often used for the transition from the convolutional layer to the fully connected layer; the fully connected layer connects each node to all nodes in the previous layer, maps the extracted features comprehensively into a one-dimensional feature vector, and this feature vector contains all feature information and can be transformed into the probability of finally classifying into each category; the SoftMax operation maps the one-dimensional class probability vector output by the previous node to the range between [0, 1] so that the sum of the probabilities is 1.

[0036] Use the cross-entropy loss to fine-tune the classification and recognition network, so as to achieve radar image target recognition. The loss function can be expressed by the following formula:

[0037]

[0038] where, y c represents the true one-hot label, and f c (x, θ) represents the feature mapping function of the classifier to obtain a vector containing the probability of each target category.

[0039] In this embodiment, the labeled image information of the radar dataset MSTAR is used in the network fine-tuning stage, and the pre-trained model of the optical image dataset ImageNet is used in the pre-training stage. A total of five classical self-supervised models (BYOL, DenseCL, MoCoV2, SimCLR, SimSiam) are used for pre-training, and their final target recognition accuracy is as Figure 2 shown. Table 2 lists the comparison results of the recognition accuracy between the method of the present invention and other few-shot radar recognition methods.

[0040] Table 2

[0041]

[0042] The experimental results show that, compared with other target recognition methods, the method of the present invention can effectively improve the accuracy and robustness of the few-shot radar target recognition network. Although ImageNet is an optical image different from radar images, due to its large amount of data, the pre-trained self-supervised model can effectively maximize the similarity features between the same-class targets and weaken the similarity features between different-class targets. In the subsequent classification network fine-tuning stage, accurate recognition of radar targets can be achieved by inputting a small amount of radar labeled data.

[0043] In summary, the method of the present invention can effectively identify target features and recognize target categories, solve the problem of insufficient network learning caused by the scarcity of radar image sample labels, improve the recognition performance of the network for radar targets under a small number of labeled data samples, and improve the accuracy and robustness of the few-shot radar target recognition network compared with other target recognition methods.

[0044] Those of ordinary skill in the art will realize that the embodiments described herein are for helping the reader understand the principles of the present invention and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations without departing from the essence of the present invention based on the technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A radar target recognition method based on self-supervised model transfer learning, the specific steps are as follows: S1. Pre-train the self-supervised model to improve the effectiveness of feature extraction; In the self-supervised model, the backbone network is pre-trained on a large number of unlabeled image samples in a self-supervised manner, and the network weights are updated using the similarity loss to maximize the similar features of the same class of targets and weaken the similar features of different classes of targets, so as to effectively extract the similar features of the same class of targets by the network; Among them, The loss used for self-supervised learning in the pre-training stage is called the similarity loss; When a pair of images is input into the self-supervised model, if they are similar, the model infers them as 1, otherwise as 0; S2. Fine-tune the classification and recognition network to achieve radar image target recognition; Transfer the network parameters of the backbone network in the pre-trained self-supervised model, add a classifier behind it, input the radar target image sample x with labels, and assume there are N c types of radar targets, then use an N c -dimensional one-hot vector y to represent the class label; the output of this classification and recognition network is the conditional probability distribution of the class label, which can be expressed by the following formula: Among them, θ represents a series of updatable parameters including the weight w and the bias b; the added classifier includes operations such as average pooling, Flatten, fully connected layer, and SoftMax; The classification and recognition network is fine-tuned using the cross-entropy loss to achieve radar image target recognition, and the loss function can be expressed by the following formula: Among them, y c represents the true value one-hot label, and f c (x, θ) represents the feature mapping function of the classifier.

Images