A lightweight few-shot radar target recognition method based on recursive design

Abstract

The present invention discloses a lightweight few-shot radar target recognition method based on recursive design. First, in the training stage, a small number of labeled samples are input into the network. Its feature maps pass through a lightweight multi-level cross-fusion CNN composed of depthwise separable convolution and coordinated attention mechanism. The loss of the distance between clusters is used to construct a feature embedding space and train the feature embedding network. Then, LM-SoftMax is used to enhance the classifier and train the last few convolutional layers. Finally, the network is updated alternately with the combination of the loss of the distance between clusters and the loss of LM-SoftMax to further optimize the network, thereby realizing target recognition. The method of the present invention can effectively utilize the method of double-loss alternating optimization to learn effective features between and within classes, and achieve accurate radar target recognition in the few-shot background. Compared with other few-shot target recognition methods, it can effectively improve the network recognition accuracy and robustness in the few-shot case.

Description

Technical Field

[0001] The present invention belongs to the field of radar image target recognition, and particularly relates to a lightweight few-shot radar target recognition method based on recursive design. 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, radar images contain the shape and scattering information of the target object and can well represent the target feature information. Synthetic aperture radar (SAR) is an important microwave remote sensing system, which has the advantages of collecting all-weather and high-resolution images day and night. SAR automatic target recognition (ATR) technology is one of the very crucial challenges in SAR applications. Therefore, the application of SAR ATR in the civilian field has important significance.

[0003] Due to its powerful feature learning ability, the Convolutional Neural Network (CNN) has made certain progress in the field of SAR ATR. However, such methods require a large number 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 the model parameters. However, its recognition performance decreases significantly with the reduction of the 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, it is crucial to improve the performance of SAR ATR through few-shot training and learning. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes a lightweight few-shot radar target recognition method based on recursive design.

[0005] The technical solution of the present invention is as follows: A lightweight few-shot radar target recognition method based on recursive design, and the specific steps are as follows:

[0006] Step 1: In the training phase, a small number of labeled samples are input into the network, and their feature maps pass through a lightweight multi-level cross-fusion CNN composed of depthwise separable convolution and coordinated attention mechanism;

[0007] Step 2: Construct a feature embedding space and train a feature embedding network using the loss of the distance between clusters;

[0008] Step 3: Use LM-SoftMax to enhance the classifier and train the last few convolutional layers;

[0009] Step 4: Update the network alternately with the combined loss of the distance between clusters and the loss of LM-SoftMax to further optimize the network, thereby achieving target recognition.

[0010] Advantages of the present invention: The method of the present invention first inputs a small number of labeled samples into the network in the training phase. Their feature maps pass through a lightweight multi-level cross-fusion CNN composed of depthwise separable convolution and coordinated attention mechanism. Use the loss of the distance between clusters to construct a feature embedding space and train a feature embedding network. Then use LM-SoftMax to enhance the classifier and train the last few convolutional layers. Finally, update the network alternately with the combined loss of the distance between clusters and the loss of LM-SoftMax to further optimize the network, thereby achieving target recognition. The method of the present invention can effectively use the method of alternating optimization of double losses to learn and identify effective features between and within categories, and achieve accurate radar target recognition in the small sample background. Compared with other small sample target recognition methods, it can effectively improve the network recognition accuracy and robustness in the case of small samples. Description of the Drawings

[0011] Figure 1 It is a flowchart of a lightweight small sample radar target recognition method based on recursive design of the present invention.

[0012] Figure 2 It is a schematic diagram of the lightweight multi-level cross-fusion CNN in the embodiment of the present invention.

[0013] Figure 3 It is an operation diagram of the coordinated attention mechanism in the network in the embodiment of the present invention.

[0014] Figure 4 It is an operation diagram of the depthwise separable convolution in the network in the embodiment of the present invention. Detailed Embodiments

[0015] The technical solution of the present invention will be further described below with reference to the drawings.

[0016] As Figure 1 shown, it is a flowchart of a lightweight small sample radar target recognition method based on recursive design of the present invention. The specific steps are as follows:

[0017] Step 1. In the training phase, a small number of labeled samples are input into the network, and their feature maps pass through a lightweight multi-level cross-fusion CNN composed of depthwise separable convolution and coordinated attention mechanism.

[0018] In limited small-sample training, the depth and width of the network face a trade-off between feature effectiveness and the number of parameters. In this embodiment, a lightweight multi-level cross-fusion CNN is adopted. As Figure 2 shown, the lightweight extractor is a bottom-up approach, which has more feature effectiveness and fewer parameters compared with ordinary convolutional layers. The structure of the lightweight extractor is constructed by stacking depthwise separable convolutions. For the features extracted by the CNN, the key features of different category targets are distributed in different levels of these features. The multi-level cross-fusion adaptively fuses the multi-level features in pairs to improve the effective input of the classifier and reduce the number of parameters of the subsequent network, which can significantly improve the feature density of the label recognition classifier in small-sample learning. Promoting feature effectiveness in a lightweight structure seems to be a paradox. The coordinated attention mechanism is adopted to capture long-range spatial interactions and perform depthwise separable convolution with fewer parameters.

[0019] The coordinated attention mechanism can promote the effectiveness of the features extracted by each layer, enable the network to initially focus on important target features, and at the same time suppress unnecessary features. The channel and position information are encoded in two steps:

[0020] As Figure 3 shown, first, the coordinated attention decomposes the global set into one-dimensional feature encoding operations, then concatenates these features in two spatial directions and sends them to a shared 1×1 convolution transformation function with batch normalization and non-linear activation functions, as well as two separate 1×1 convolution transformation functions to more accurately locate and identify the objects of interest.

[0021] The depthwise separable convolution can decouple the correlation between the channel and spatial dimensions in the convolutional layer and map them separately, so as to achieve better results with fewer parameters than ordinary convolution. This means that the depthwise separable convolution can work together with the coordinated attention mechanism to further promote the effectiveness of the extracted features. As Figure 4 shown, the depthwise separable convolution has two steps: depthwise separable convolution and pointwise separable convolution.

[0022] Step 2. The loss of the distance between clusters is used to construct the feature embedding space and train the feature embedding network.

[0023] In this embodiment, the purpose of the second step is to utilize the distinction of features to strengthen the aggregation of intra-class features and the separation of inter-class features, and introduce the concept of mutual features, which refers to the common features among SAR images of the same target type. For a classification and recognition task with N samples and K categories, the mutual features m of the k-th type of target are extracted from the flattened feature map using principal component analysis (PCA) technology. k , for all samples {x1, x2,..., x i} in the k-th type of target, the Euclidean distance is used to calculate the distance between one sample x i and the mutual feature m k , which can be expressed as:

[0024] d(x i , m k ) = ||ζ(x i ) - m k || 2 (1)

[0025] Among them, ||·|| 2 represents the L2 norm operation, and ξ(·) represents the output of the flattened feature map after PCA operation.

[0026] The Euclidean distance between the mutual features of different target categories m k and m l can be calculated by Equation (2):

[0027] d′(m k , m l ) = -||m k - m l || 2 (2)

[0028] Among them, m l refers to the mutual features of the l-th type of target.

[0029] Then, the loss used to identify the inter-cluster distance of features can be calculated by Equation (3):

[0030]

[0031] Among them, m i refers to the mutual features of the i-th type of target, refers to the i-th sample in the k-th type of target. The first summation in Equation (3) forces each sample of the k-th type of target to be close to the mutual feature m k , realizing intra-class feature aggregation; the second summation in Equation (3) aims to separate the mutual features of different target types, realizing inter-class feature separation.

[0032] Step 3: LM-SoftMax is used to enhance the classifier and train the last few convolutional layers.

[0033] Compared with the standard SoftMax, LM-SoftMax can greatly enhance the intra-class feature aggregation and inter-class feature separation based on feature discrimination. Consider a classification and recognition task with N samples and K categories. Assume that the output vector of the network before inputting the standard SoftMax is f i is the i-th element of the vector , and the input sample to the network is x i . Then the probability of predicting the sample x i as the category y i can be expressed as:

[0034]

[0035] where k ∈ [1, K] is the sample label, and the loss function of the standard SoftMax can be calculated by Equation (5):

[0036]

[0037] When the category prediction of the sample x i is correct, for k ∈ [1, K], the meaning of Equation (5) can be expressed by Equation (6):

[0038]

[0039] where, refers to the intermediate embedding of the input sample x i , and W k refer to the weight vectors of multiple layers of convolution, and the superscript T is the matrix transpose.

[0040] LM-SoftMax requires where, refers to the angle between k and x, θ k is the angle between W

[0041] and x, and m is a pre-set positive integer. When the value of m increases, the margin between categories will become larger, and the training and learning will become difficult.

[0042]

[0043] where, ψ(θ) = (-1) qcos(mθ)-2q, where θ ∈ [(qπ / m), ((q + 1)π / m)] and q is a positive integer in the range of [0, m - 1]. Different from the standard SoftMax calculation in Equation (5), LM-SoftMax can be calculated by Equation (8):

[0044]

[0045] where y k indicates that the predicted label is the k-th class.

[0046] The loss function of LM-SoftMax can be expressed by Equation (9):

[0047]

[0048] Step 4: Update the network by alternately combining the loss of the distance between clusters and the loss of LM-SoftMax to further optimize the network, thereby achieving target recognition.

[0049] Train the entire network by gradually controlling the proportion of the two losses in each training step. At the t-th training step, if the loss of LM-SoftMax is the dominant loss, then the total loss at the t-th step can be expressed as:

[0050]

[0051] where α is a parameter for adjusting the proportion of the loss of the distance between clusters in the total loss at the (t - 1)-th training step and is the loss of LM-SoftMax. At the first training step t = 1, the total loss is initialized as If the loss of the distance between clusters is the dominant loss, then the total loss at the t-th step can be expressed as:

[0052]

[0053] where β is a parameter for adjusting the proportion of the loss of LM-SoftMax in the total loss at the (t - 1)-th training step and

[0054] By updating the network with such a combined alternating double loss, both the embedding network and the classifier can be more effectively optimized. During the training process, the integration of the distance between clusters and the LM-SoftMax loss can help optimize the entire model to achieve better collaboration and higher recognition accuracy, thereby realizing the SAR target recognition technology under small samples.

[0055] In this embodiment, compared with other small-sample object recognition methods, the method of the present invention can effectively improve the network recognition accuracy and robustness in the case of small samples. The comparison data of the recognition accuracy rates with the remaining small-sample methods are shown in Table 1 as follows:

[0056] Table 1

[0057]

Claims

1. A lightweight few-shot radar target recognition method based on recursive design, the specific steps are as follows: Step 1: In the training stage, a small number of labeled samples are input into the network, and their feature maps pass through a lightweight multi-level cross-fusion CNN composed of depthwise separable convolution and coordinated attention mechanism; Step 2: Use the loss of the distance between clusters to construct a feature embedding space and train a feature embedding network; Step 3: Use LM-SoftMax to improve the classifier and train the last few convolutional layers; In the said Step 3: Set up a classification and recognition task with N samples and K categories. Assume that the output vector of the network before the input standard SoftMax is f i is the i-th element of the vector and the sample input to the network is x i . Then the probability of predicting the sample x i as the category y i is expressed as: Among them, k ∈ [1, K] is the sample label, and the loss function of the standard SoftMax can be calculated by Equation (2): When the sample x i is correctly predicted in terms of its class, for k ∈ [1, K], the meaning of Equation (2) can be represented by Equation (3): Among them, refers to the intermediate embedding of the input sample x i , and W k refers to the weight vector of the multi-layer convolution, and the superscript T is the matrix transpose; LM-SoftMax requires Among them, refers to the angle between and x, θ k is the angle between W k and x, and m is a pre-set positive integer; LM-SoftMax requires strict inequality constraints, as shown in Equation (4): where ψ(θ) = (-1) q cos(mθ) - 2q, and θ ∈ [(qπ / m), ((q + 1)π / m)], q is a positive integer in the range [0, m - 1], and LM-SoftMax can be calculated by Equation (5): where y k indicates that the predicted label is the k-th class; The loss function of LM-SoftMax can be expressed by Equation (6): Step 4: Update the network by alternately combining the loss of the distance between clusters and the loss of LM-SoftMax to further optimize the network, so as to achieve target recognition.

2. A lightweight small-sample radar target recognition method based on recursive design according to claim 1, characterized in that, In the said Step 2: Suppose there is a classification and recognition task with N samples and K categories. The mutual feature m of the k-th category of objects is extracted from the flattened feature map using the principal component analysis technique. k , for all samples {x1, x2,..., x i} in the k-th category of objects, the Euclidean distance is used to calculate the distance between one sample x i and the mutual feature m k , which can be expressed as: d(x i ,m k ) = ||ζ(x i ) - m k || 2 (7) where, ||·|| 2 represents the L2 norm operation, and ξ(·) represents the output of the flattened feature map after the PCA operation; Different target categories m k and m l The Euclidean distance between the mutual features can be calculated by Equation (8): d′(m k ,m l )=-||m k -m l || 2 (8) where m l refers to the mutual features of the l-th type of target; Then the loss of the distance between clusters for identifying features can be calculated by Equation (9): where m i represents the mutual features of the i-th type of target, and represents the i-th sample in the k-th type of target.

3. A lightweight few-shot radar target recognition method based on recursive design according to claim 1, characterized in that In the said Step 4: Train the entire network by gradually controlling the proportion of the two losses in each step of training. At the t-th step of training, if the loss of LM-SoftMax is the dominant loss, then the total loss at the t-th step can be expressed as: Among them, α is a parameter that adjusts the proportion of the loss of the distance between clusters in the training at the (t - 1) step in the total loss The parameter of the proportion of the total loss is the loss of LM-SoftMax. At the first step of training t = 1, the total loss is initialized as If the loss of the distance between clusters is the dominant loss, the total loss at the t-th step is expressed as: Among them, β is a parameter for adjusting the proportion of the loss of LM-SoftMax in the (t-1)-th step of training to the total loss. ​

Images