A Segmentation-Assisted Semi-Supervised Automatic Target Recognition Method for Radar Images

Abstract

The present invention discloses a method for automatic target recognition of semi-supervised radar images with auxiliary segmentation. First, in the training stage, the processed training data is input into the network, and two semi-blind trainings based on recognition and segmentation are sequentially performed to form a training cycle process (TCL). Based on the proposed IRL, the information of recognition loss and segmentation loss in the semi-blind training is combined to update the corresponding parameter settings of the network, and the training cycle is continuously repeated. The migration information of unlabeled samples is used to construct an effective inductive bias on a small number of labeled samples to optimize the network settings and achieve accurate recognition of the target. The method of the present invention adopts a training cycle process, applies the auxiliary segmentation information of unlabeled samples to construct an inductive bias, achieves high-performance target recognition under a small number of labeled samples, provides a new information source for SAR ATR, breaks through the limited information limit of a small number of labeled sample training, achieves high-performance SAR ATR with good robustness, and shows excellent performance in the MSTAR data set.

Description

Technical Field

[0001] The invention belongs to the field of SAR image target recognition, and in particular relates to an auxiliary segmentation semi-supervised radar image automatic target recognition method. Background Art

[0002] Synthetic aperture radar (SAR) is an important microwave remote sensing system in the civilian field. SAR automatic target recognition (ATR) is one of the most critical and challenging issues in SAR applications. In some important application scenarios, such as earthquake rescue and maritime rescue, the number of available SAR images may be very small. Therefore, the study of SAR ATR methods under the limitation of a small number of samples has important practical application significance.

[0003] Existing Few-Shot Learning (FSL) methods in SAR ATR can be mainly divided into three categories: data augmentation methods, metric-based methods, and model-based methods. The paper "Zheng, Ce, Xue Jiang, and Xingzhao Liu, Semi-supervised SAR ATR via multi-discriminator generative adversarial network, IEEE Sensors Journal, vol.19, no.17, pp.7525-7533, 2019" proposes a semi-supervised SAR ATR method via multi-discriminator generative adversarial network (GAN), which belongs to data augmentation methods and achieves high performance by enriching the number of training SAR images and improving training quality. The paper "Zheng, Ce, Xue Jiang, and Xingzhao Liu, Semi-supervised SAR ATR via multi-discriminator generative adversarial network, IEEE Sensors Journal, vol.19, no.17, pp.7525-7533, 2019" proposes a conv-biLSTM prototype network, which is a metric-based method that learns representations based on categories that can be generalized to new categories. The paper "Persello, Claudio, and Lorenzo Bruzzone, Active and semisupervised learning for the classification of remote sensing images, IEEE Transactions on Geoscience and Remote Sensing, vol.52, no.11, pp.6937-6956, 2014" proposes an active and semi-supervised learning network for SAR ATR, which is a semi-supervised support vector machine (SVM), a model-based method that uses prior knowledge to construct an embedding space and constrain complexity.

[0004] The key challenge of FSL in SAR ATR lies in whether there is a useful inductive bias. The above methods are essentially just continuously mining a few labeled SAR images to obtain more usable recognition information. In other words, from the perspective of information theory, these methods cannot obtain new information from samples, so the SAR image recognition performance is very limited. Summary of the invention

[0005] To solve the above technical problems, the present invention proposes a semi-supervised automatic target recognition method for radar images with assisted segmentation. To simplify the expression, in the following text, samples with class labels but no segmentation labels are referred to as labeled samples, and samples with segmentation labels but no class labels are referred to as unlabeled samples.

[0006] The technical solution of the present invention is: a semi-supervised radar image automatic target recognition method with auxiliary segmentation, and the specific steps are as follows:

[0007] Step 1: In the training phase, the processed training data, including the data sets of unlabeled samples and labeled samples required for training, are input into the network together;

[0008] Step 2: The training data undergoes two semi-blind trainings based on recognition and segmentation respectively to form a training cycle process (TCL);

[0009] Step 3: Based on the proposed IRL, combined with the information of recognition loss and segmentation loss in semi-blind training, update the corresponding network parameter settings;

[0010] Step 4: By continuously repeating the above training cycle, the migration information of unlabeled samples is gradually utilized to build effective inductive biases on a few labeled samples to optimize the network settings, thereby achieving accurate recognition of the target.

[0011] Beneficial effects of the present invention: The method of the present invention first inputs the processed training data into the network during the training phase, and sequentially undergoes two semi-blind trainings based on recognition and segmentation to form a training cycle process (TCL). Based on the proposed IRL, combined with the information of recognition loss and segmentation loss in the semi-blind training, the corresponding parameter settings of the network are updated, and the training cycle is continuously repeated. The migration information of unlabeled samples is used to construct an effective inductive bias on a small number of labeled samples to optimize the network settings and achieve accurate recognition of the target. The method of the present invention adopts a training cycle process, applies the auxiliary segmentation information of unlabeled samples to construct an inductive bias, achieves high-performance target recognition under a small number of labeled samples, provides a new source of information for SAR ATR, breaks through the limited information limit of a small number of labeled sample training, achieves high-performance SAR ATR with good robustness, and demonstrates excellent performance in the MSTAR data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The present invention is a flow chart of a method for automatic target recognition in semi-supervised radar images with auxiliary segmentation.

[0013] Figure 2 The present invention is a schematic diagram of the framework and benchmark structure of a semi-supervised radar image automatic target recognition method with auxiliary segmentation. DETAILED DESCRIPTION

[0014] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0015] like Figure 1 As shown, a flow chart of a method for automatic target recognition of semi-supervised radar images with auxiliary segmentation of the present invention is shown, and the specific steps are as follows:

[0016] Step 1: In the training phase, the processed training data, including the unlabeled set D required for training, is U (samples with segmentation labels but no class labels) and the label set D L (samples with class labels but no segmentation labels) are input into the network for training.

[0017] The network consists of a shared feature extractor, a classifier for labeled SAR image recognition, and a decoder for unlabeled SAR image segmentation. The extractor, decoder, and classifier are represented as M E 、M D 、M C , and denote the corresponding learnable parameter as P E , P D , P C .

[0018] Shared feature extractor M E ,like Figure 2 As shown in Figure 1. The convolutional layers are connected in a residual manner, and the attention mechanism block containing the spatial and channel attention modules is used to focus on important local features and compress useless information at each layer. In addition, leaky ReLu (rectified linear unit with leakage) and batch normalization are used to improve its generalization ability.

[0019] Decoder M for unlabeled SAR image segmentation D , which consists of a deconvolution module and a segmentation layer, such as Figure 2 As shown in the figure, the deconvolution module consists of two parts: a cross-convolution layer for reconstructing the SAR image structure, and a convolution layer with attention and batch normalization structure for enhancing generalization ability. The segmentation layer (Conv3-2) is arranged at the end of these deconvolution modules and outputs the segmentation result.

[0020] Classifier M for labeled SAR image recognition C ,like Figure 2 It controls the parameters of the classifier by simply stacking convolutional layers and maximum pooling layers, which is an effective structure of the classifier in SAR ATR.

[0021] Step 2: The training data undergoes two semi-blind trainings based on recognition and segmentation to form a training cycle (TCL).

[0022] First, the unlabeled set D U Sequentially through the shared feature extractor M E and decoder M D , to achieve a segmentation-based semi-blind training. Then, a small number of labeled sets D L Sequentially through the shared feature extractor M E and classifier M C , achieving a semi-blind training based on recognition.

[0023] Step 3: Based on the proposed IRL, combined with the information of recognition loss and segmentation loss in semi-blind training, update the corresponding network parameter settings.

[0024] During training, first, the segmentation loss L is calculated by semi-blind training based on segmentation. S Used to update the decoder corresponding parameters P D , L S The function is defined as:

[0025]

[0026] Among them, m(s i |x) represents the binary matrix of the segmentation results of all pixels on the i-th SAR slice image, x represents the input SAR image, n represents the number of pixels in the SAR slice image, s i represents the segmentation label, and V represents the number of segmentation types.

[0027] Then, the recognition loss L is calculated by semi-blind training based on recognition R Used to update the corresponding parameters P of the classifier C , L R The function is defined as:

[0028]

[0029] Among them, p(y i |x) is the probability vector of the recognition result on the i-th SAR slice image, y i is the identification label, and C is the number of identification classes.

[0030] Meanwhile, in the training loop, after the segmentation-based semi-blind training, L S and the recognition loss of the previous training cycle based on IRL, which is used to update the parameters P of the shared feature extractor E ; After semi-blind training based on recognition, L R And the IRL-based segmentation loss, again update the parameters P of the shared feature extractor E .

[0031] Among them, the recognition loss based on IRL The definition is as follows:

[0032]

[0033] Among them, t is the number of training times, and it is used as a superscript to represent the corresponding result of the t-th training, that is, represents the identification label of the t-th training, Represents the probability vector of the recognition result on the i-th SAR slice image of the t-th training. And so on, represents the segmentation label of the t-1th training, The binary matrix represents the segmentation results of all pixels on the i-th SAR slice image of the t-1th training. α represents the parameter for adjusting the residual rate of the segmentation information of the last cycle, which is calculated as:

[0034]

[0035] IRL-based segmentation loss The definition is as follows:

[0036]

[0037] Step 4: By continuously repeating the above training cycle, the migration information of unlabeled samples is gradually utilized to build effective inductive biases on a few labeled samples to optimize the network settings, thereby achieving accurate recognition of the target.

[0038] During training, by utilizing the transfer information of unlabeled samples and combining the recognition and segmentation losses, the feature extractor is continuously optimized towards higher recognition and segmentation accuracy. At the same time, the network is forced to gradually build effective inductive biases on a few labeled samples and extract more general and effective features from unlabeled and labeled samples, ultimately achieving accurate recognition of the target.

[0039] In this embodiment, compared with other small sample target recognition methods, the method of the present invention uses auxiliary segmentation information of unlabeled samples and designed information residual loss (IRL), adopts a training cycle process, and constructs an effective inductive bias, thereby achieving high-performance target recognition with a small number of labeled samples. Experiments have shown that the present invention exhibits excellent performance in the MSTAR dataset, and the recognition accuracy comparison data with other small sample methods are shown in Table 1:

[0040] Table 1

[0041]

Claims

1. A semi-supervised radar image automatic target recognition method with auxiliary segmentation, the specific steps are as follows: Step 1: In the training phase, the processed training data, including the data sets of unlabeled samples and labeled samples required for training, are input into the network together; in, Samples with class labels but no segmentation labels are called labeled samples, and samples with segmentation labels but no class labels are called unlabeled samples. Step 2: The training data undergoes two semi-blind trainings based on recognition and segmentation respectively to form a training cycle process TCL; Step 3: Based on the proposed IRL, combined with the information of recognition loss and segmentation loss in semi-blind training, update the corresponding network parameter settings; Step 4: By continuously repeating the above training cycle, the migration information of unlabeled samples is gradually utilized to build effective inductive biases on a few labeled samples to optimize the network settings, thereby achieving accurate recognition of the target; In the step 1: The network consists of a shared feature extractor, a classifier for labeled SAR image recognition, and a decoder for unlabeled SAR image segmentation. The extractor, decoder, and classifier are represented as M E 、M D 、M C , and denote the corresponding learnable parameter as P E , P D , P C ; In the step 2: The unlabeled set D U Sequentially through the shared feature extractor M E and decoder M D , implement a segmentation-based semi-blind training, and then use a small set of labeled L Sequentially through the shared feature extractor M E and classifier M C , to achieve a semi-blind training based on recognition; In the step three: The segmentation loss L is calculated during training by segmentation-based semi-blind training. S Used to update the decoder corresponding parameters P D , L S The function is defined as: Among them, m(s i |x) represents the binary matrix of the segmentation results of all pixels on the i-th SAR slice image, x represents the input SAR image, n represents the number of pixels in the SAR slice image, s i represents the segmentation label, and V represents the number of segmentation types; Calculate the recognition loss L through recognition-based semi-blind training R Used to update the corresponding parameters P of the classifier C , L R The function is defined as: Among them, p(y i |x) is the probability vector of the recognition result on the i-th SAR slice image, y i is the identification label, C is the number of identification classes; In the training loop, after the segmentation-based semi-blind training, L S and the recognition loss of the previous training cycle based on IRL, which is used to update the parameters P of the shared feature extractor E ; After semi-blind training based on recognition, L R And the IRL-based segmentation loss, again update the parameters P of the shared feature extractor E ; Among them, the recognition loss based on IRL The definition is as follows: Among them, t is the number of training times, and it is used as a superscript to represent the corresponding result of the t-th training, that is, represents the identification label of the t-th training, represents the probability vector of the recognition result on the i-th SAR slice image of the t-th training, and so on. represents the segmentation label of the t-1th training, represents the binary matrix of the segmentation results of all pixels on the i-th SAR slice image of the t-1th training, and α represents the parameter for adjusting the residual rate of the segmentation information of the last cycle, which is calculated as: IRL-based segmentation loss The definition is as follows:

2. The method for automatic target recognition in semi-supervised radar images with auxiliary segmentation according to claim 1, characterized in that: In the step 4: During training, the migration information of unlabeled samples is used, combined with the recognition and segmentation losses, to continuously optimize the feature extractor towards higher recognition and segmentation accuracy. At the same time, the network is forced to gradually build effective inductive biases on a few labeled samples and extract more general and effective features from unlabeled and labeled samples, ultimately achieving accurate recognition of the target.

Images