Limited label data SAR (Synthetic Aperture Radar) target classification method and system based on deep convolutional neural network transfer learning

Abstract

The invention discloses a limited label data SAR target classification method and system based on deep convolutional neural network transfer learning, and belongs to the technical field of synthetic aperture radar recognition. The method comprises the following steps: constructing a source domain model by using a stacked convolution automatic encoder; processing the SAR scene images of different unmarked areas to obtain a source domain data set; performing unsupervised training on the source domain model by using the source domain data set; using the source domain model after unsupervised training to construct a target domain model; carrying out transfer learning on the target domain model by utilizing the expanded target domain data set with the mark so as to carry out fine tuning on the target domain model; and carrying out SAR target classification and identification by adopting the fine-tuned target domain model. According to the method, the problem that the scale of the SAR target data set is small is solved, and the accuracy and robustness of SAR target classification are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of synthetic aperture radar recognition, and in particular relates to a limited label data SAR target classification method and system based on deep convolutional neural network transfer learning. Background Art

[0002] Synthetic aperture radar (SAR) automatic target recognition (SAR-ATR) has important application value in military and civilian fields. Traditional SAR target classification methods are mainly based on manual feature extraction and classifier design, which are difficult to adapt to complex scenes and target changes. Deep convolutional neural networks (CNNs) have achieved great success in the field of image recognition, but their application in SAR target classification is limited. The main reason is that the scale of SAR target datasets is small, it is difficult to train large CNN models, and the accuracy of SAR target classification is low. Summary of the invention

[0003] One of the purposes of the present invention is to provide a SAR target classification method with limited label data based on deep convolutional neural network transfer learning, which solves the problem of small size of SAR target data set and improves the accuracy and robustness of SAR target classification.

[0004] The second object of the present invention is to provide a SAR target classification system with limited label data based on deep convolutional neural network transfer learning.

[0005] In order to achieve one of the above purposes, the present invention adopts the following technical solution: A method for SAR target classification with limited label data based on deep convolutional neural network transfer learning, the method comprising: Step S1: Use stacked convolutional autoencoders to build a source domain model; Step S2, processing unlabeled SAR scene images of different regions to obtain a source domain data set; Step S3: using the source domain dataset to perform unsupervised training on the source domain model; Step S4: construct a target domain model using the source domain model after unsupervised training; Step S5: using the expanded labeled target domain data set to perform transfer learning on the target domain model to fine-tune the target domain model; Step S6: Use the fine-tuned target domain model to perform SAR target classification and recognition.

[0006] Further, in the step S1, the source domain model includes an encoding part and a decoding part of a stacked convolutional autoencoder; The encoding part includes a plurality of convolutional layers with increasing number of feature map channels; each convolutional layer is connected in turn with an activation function and a The maximum pooling layer of The decoding part includes a deconvolution layer corresponding to each convolution layer; each deconvolution layer is used for upsampling; and each deconvolution layer is connected with an activation function.

[0007] Furthermore, in step S2, the area includes an urban area, a forest area, a mountain area and a cultivated land area.

[0008] Furthermore, in step S4, the target domain model is constructed as follows: Step S41, adding two fully connected layers and a classifier after the encoding part in the source domain model after unsupervised training to form a classification path of the target domain model; Step S42, adding 5 deconvolution layers after the decoding part in the source domain model after unsupervised training to form a reconstruction path of the target domain model; Step S43: forming the target domain model by using the classification path and reconstruction path of the target domain model.

[0009] Furthermore, in step S5, the specific process of the fine-tuning is: The decoding part in the reconstruction path is frozen, and the encoding part in the classification path is fine-tuned.

[0010] In order to achieve the second objective above, the present invention adopts the following technical solution: A limited label data SAR target classification system based on deep convolutional neural network transfer learning, the limited label data SAR target classification system comprising: The first building module is used to build a source domain model using a stacked convolutional autoencoder; A processing module, used for processing unlabeled SAR scene images of different regions to obtain a source domain data set; An unsupervised training module, used to perform unsupervised training on the source domain model using the source domain dataset; The second construction module is used to construct a target domain model using the source domain model after unsupervised training; A transfer learning module, used to perform transfer learning on the target domain model using the expanded labeled target domain dataset to fine-tune the target domain model; The classification and recognition module is used to perform SAR target classification and recognition using the fine-tuned target domain model.

[0011] Further, the source domain model includes an encoding part and a decoding part of a stacked convolutional autoencoder; The encoding part includes a plurality of convolutional layers with increasing number of feature map channels; each convolutional layer is connected in turn with an activation function and a The maximum pooling layer of The decoding part includes a deconvolution layer corresponding to each convolution layer; each deconvolution layer is used for upsampling; and each deconvolution layer is connected with an activation function.

[0012] Furthermore, the areas include urban areas, forest areas, mountainous areas and cultivated land areas.

[0013] Furthermore, the second building block includes: A first adding submodule is used to add two fully connected layers and a classifier after the encoding part in the source domain model after unsupervised training to form a classification path of the target domain model; A second adding submodule is used to add 5 deconvolution layers after the decoding part in the source domain model after unsupervised training to form a reconstruction path of the target domain model; The forming submodule is used to form the target domain model by using the classification path and the reconstruction path of the target domain model.

[0014] Furthermore, the fine-tuning in the transfer learning module performs the following steps: The decoding part in the reconstruction path is frozen to fine-tune the encoding part in the classification path.

[0015] In summary, the technical solution of the present invention has the following technical effects: The present invention uses unlabeled SAR scene images of different regions (i.e., source domain data sets) to perform unsupervised training on a source domain model constructed using a stacked convolutional autoencoder to learn the hierarchical feature representation of the SAR scene images; and uses the source domain model after the unsupervised training to construct a target domain model; uses an expanded labeled target domain data set to perform transfer learning on the target domain model to fine-tune the target domain model; and uses the fine-tuned target domain model to perform SAR target classification and recognition, thereby solving the problem that the SAR target data set is small in scale and it is difficult to train a large CNN model, and improving the accuracy and robustness of SAR target classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 It is a flow chart of the method for SAR target classification with limited label data based on deep convolutional neural network transfer learning of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0019] This embodiment provides a limited label data SAR target classification method based on deep convolutional neural network transfer learning, refer to Figure 1 , the limited label data SAR target classification method includes: Step S1: Use stacked convolutional autoencoders to build a source domain model.

[0020] The encoding part of the stacked convolutional autoencoder SCAE in this embodiment includes multiple convolutional layers (such as 5 convolutional layers) and pooling layers (including ReLU activation function and The maximum pooling layer of the convolution layer adopts a pyramid structure, and the number of feature map channels increases successively. Each convolution layer is connected to a ReLU activation function and a The decoding part is symmetrical with the encoding part, and uses a deconvolution layer for upsampling and a ReLU activation function. The mean square error (MSE) is used as the loss function during training, and the mini-batch stochastic gradient descent (SGD) algorithm is used for optimization.

[0021] Based on the above concept, the source domain model of this embodiment includes an encoding part and a decoding part of a stacked convolutional autoencoder; The encoding part includes a plurality of convolutional layers with increasing number of feature map channels; each convolutional layer is connected in turn with an activation function and a The maximum pooling layer of The decoding part includes a deconvolution layer corresponding to each convolution layer; each deconvolution layer is used for upsampling; and each deconvolution layer is connected with an activation function.

[0022] Step S2: Process the unlabeled SAR scene images of different regions to obtain a source domain data set.

[0023] The source domain dataset is constructed from a large number of unlabeled SAR scene images collected by the German earth observation satellite TerraSAR-X, covering various landscape areas such as cities, forests, mountains, and cultivated land (that is, the area includes urban areas, forest areas, mountain areas, and cultivated land areas.), and is cropped into Pixels are collected and normalized to obtain the source domain dataset.

[0024] Step S3: Using the source domain dataset, perform unsupervised training on the source domain model.

[0025] This embodiment uses a stacked convolutional autoencoder (SCAE) to perform unsupervised training on a source domain dataset to learn the hierarchical feature representation of SAR scene images and obtain a source domain model after unsupervised training.

[0026] The mean square error (MSE) is used as the loss function during training, and the mini-batch stochastic gradient descent (SGD) algorithm is adopted for optimization.

[0027] Step S4: construct a target domain model using the source domain model after unsupervised training.

[0028] The target domain model of this embodiment includes a classification path and a reconstruction path. The classification path connects two fully connected layers and a Softmax classifier after the encoding part of the source domain model for SAR target classification. The reconstruction path connects five deconvolution layers after the encoding part of the source domain model for reconstructing SAR target images. The construction process of the target domain model is as follows: Step S41, adding two fully connected layers and a classifier after the encoding part in the source domain model after unsupervised training to form a classification path of the target domain model to achieve SAR target classification; Step S42, adding 5 deconvolution layers after the decoding part in the source domain model after unsupervised training to form a reconstruction path of the target domain model to achieve SAR target image reconstruction; Step S43: forming the target domain model by using the classification path and reconstruction path of the target domain model.

[0029] Step S5: Using the expanded labeled target domain data set, transfer learning is performed on the target domain model to fine-tune the target domain model.

[0030] The target domain dataset is constructed by selecting images of 10 types of vehicles from the Mobile and Stationary Target Acquisition and Recognition (MSTAR) dataset: T72, BTR70, BMP2, 2S1, BRDM2, BTR60, D7, T62, ZIL131 and ZSU23. The images contain data at two angles of 15° and 17°. The images are preprocessed by cropping and normalizing, and data augmentation is performed, such as translation and mirroring, to obtain the augmented target domain dataset with labels.

[0031] The feature representation learned in the source domain model is transferred to the target domain model and fine-tuned to adapt to the SAR target classification task. During the transfer learning process, the encoding part of the source domain model is frozen, and only the classification path and the decoding part of the reconstruction path of the target domain model are fine-tuned. The specific process of fine-tuning is: The decoding part in the reconstruction path is frozen, and the encoding part in the classification path is fine-tuned.

[0032] Step S6: Use the fine-tuned target domain model to perform SAR target classification and recognition.

[0033] The target domain model after transfer learning is used to classify and identify SAR targets.

[0034] The classification results are evaluated using a confusion matrix, and indicators such as accuracy, recall, F1 value, and mean average precision (mAP) are calculated. In order to verify the effectiveness of the method of the present invention, it is compared with the following methods: 1. SVM: Use support vector machine for SAR target classification and train it using manually extracted features.

[0035] 2. MSRC: SAR target classification using single gene signal sparse representation method.

[0036] 3. TJSR: SAR target classification using three-task joint sparse representation method.

[0037] 4. SDDLSR: SAR target classification using supervised discriminative dictionary learning and sparse representation methods.

[0038] 5. JDSR: SAR target classification using joint dynamic sparse representation method.

[0039] 6. ConvNet: Use a fully convolutional network for SAR target classification, trained from scratch.

[0040] 7. DCHUN: SAR target classification using deep convolutional neural networks and fine-tuning using pre-trained models.

[0041] Experimental results show that the classification accuracy, recall rate, F1 value and mAP of this embodiment on the MSTAR dataset are better than those of the above-mentioned comparison methods. Especially when the dataset size is small, the performance improvement of the method of the present invention is more obvious.

[0042] This embodiment uses unlabeled SAR scene images of different regions (i.e., source domain data sets) to perform unsupervised training on a source domain model constructed using a stacked convolutional autoencoder to learn the hierarchical feature representation of the SAR scene images; and uses the source domain model after unsupervised training to construct a target domain model; uses an expanded labeled target domain data set to perform transfer learning on the target domain model to fine-tune the target domain model; and uses the fine-tuned target domain model to perform SAR target classification and recognition, which solves the problem that the SAR target data set is small in scale and it is difficult to train a large CNN model, and improves the accuracy and robustness of SAR target classification.

[0043] The above embodiment can be implemented by adopting the technical solutions of the following embodiments: Another embodiment provides a limited label data SAR target classification system based on deep convolutional neural network transfer learning, and the limited label data SAR target classification system includes: The first building module is used to build a source domain model using a stacked convolutional autoencoder; A processing module, used for processing unlabeled SAR scene images of different regions to obtain a source domain data set; An unsupervised training module, used to perform unsupervised training on the source domain model using the source domain dataset; The second construction module is used to construct a target domain model using the source domain model after unsupervised training; A transfer learning module, used to perform transfer learning on the target domain model using the expanded labeled target domain dataset to fine-tune the target domain model; The classification and recognition module is used to perform SAR target classification and recognition using the fine-tuned target domain model.

[0044] Further, the source domain model includes an encoding part and a decoding part of a stacked convolutional autoencoder; The encoding part includes a plurality of convolutional layers with increasing number of feature map channels; each convolutional layer is connected in turn with an activation function and a The maximum pooling layer of The decoding part includes a deconvolution layer corresponding to each convolution layer; each deconvolution layer is used for upsampling; and each deconvolution layer is connected with an activation function.

[0045] Furthermore, the areas include urban areas, forest areas, mountainous areas and cultivated land areas.

[0046] Furthermore, the second building block includes: A first adding submodule is used to add two fully connected layers and a classifier after the encoding part in the source domain model after unsupervised training to form a classification path of the target domain model; A second adding submodule is used to add 5 deconvolution layers after the decoding part in the source domain model after unsupervised training to form a reconstruction path of the target domain model; The forming submodule is used to form the target domain model by using the classification path and the reconstruction path of the target domain model.

[0047] Furthermore, the fine-tuning in the transfer learning module performs the following steps: Freezing a decoding portion in the reconstruction path to fine-tune an encoding portion in the classification path; The principles, formulas and parameter definitions involved in the above embodiments are all applicable and will not be described in detail here.

[0048] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. The above embodiments only express several implementation methods of the present application, and their descriptions are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, without departing from the concept of the present application, several deformations and improvements can be made, which all belong to the protection scope of the present application. Therefore, the protection scope of the patent of this application shall be based on the attached claims.

Claims

1. A SAR target classification method with limited label data based on deep convolutional neural network transfer learning, characterized in that: The limited label data SAR target classification method comprises: Step S1: Use stacked convolutional autoencoders to build a source domain model; Step S2, processing unlabeled SAR scene images of different regions to obtain a source domain data set; Step S3: using the source domain dataset to perform unsupervised training on the source domain model; Step S4: construct a target domain model using the source domain model after unsupervised training; Step S5: using the expanded labeled target domain data set to perform transfer learning on the target domain model to fine-tune the target domain model; Step S6: Use the fine-tuned target domain model to perform SAR target classification and recognition.

2. The limited label data SAR target classification method according to claim 1 is characterized in that: In the step S1, the source domain model includes an encoding part and a decoding part of a stacked convolutional autoencoder; The encoding part includes a plurality of convolutional layers with increasing number of feature map channels; each convolutional layer is connected in turn with an activation function and a The maximum pooling layer of The decoding part includes a deconvolution layer corresponding to each convolution layer; each deconvolution layer is used for upsampling; and each deconvolution layer is connected with an activation function.

3. The limited label data SAR target classification method according to claim 2 is characterized in that: In the step S2, the area includes an urban area, a forest area, a mountain area and a cultivated land area.

4. The limited label data SAR target classification method according to any one of claims 1 to 3, characterized in that: In step S4, the target domain model is constructed as follows: Step S41, adding two fully connected layers and a classifier after the encoding part in the source domain model after unsupervised training to form a classification path of the target domain model; Step S42, adding 5 deconvolution layers after the decoding part in the source domain model after unsupervised training to form a reconstruction path of the target domain model; Step S43: forming the target domain model by using the classification path and reconstruction path of the target domain model.

5. The limited label data SAR target classification method according to claim 4 is characterized in that: In step S5, the specific process of the fine-tuning is: The decoding part in the reconstruction path is frozen, and the encoding part in the classification path is fine-tuned.

6. A SAR target classification system with limited label data based on deep convolutional neural network transfer learning, characterized in that: The limited label data SAR target classification system comprises: The first building module is used to build a source domain model using a stacked convolutional autoencoder; A processing module, used for processing unlabeled SAR scene images of different regions to obtain a source domain data set; An unsupervised training module, used to perform unsupervised training on the source domain model using the source domain dataset; The second construction module is used to construct a target domain model using the source domain model after unsupervised training; A transfer learning module, used to perform transfer learning on the target domain model using the expanded labeled target domain dataset to fine-tune the target domain model; The classification and recognition module is used to perform SAR target classification and recognition using the fine-tuned target domain model.

7. The limited label data SAR target classification system according to claim 6, characterized in that: The source domain model includes an encoding part and a decoding part of a stacked convolutional autoencoder; The encoding part includes a plurality of convolutional layers with increasing number of feature map channels; each convolutional layer is connected in turn with an activation function and a The decoding part includes a deconvolution layer corresponding to each convolution layer; each deconvolution layer is used for upsampling; and an activation function is connected after each deconvolution layer.

8. The limited label data SAR target classification system according to claim 7, characterized in that: The areas include urban areas, forest areas, mountainous areas and cultivated land areas.

9. The limited label data SAR target classification system according to any one of claims 6 to 8, characterized in that: The second building block comprises: A first adding submodule is used to add two fully connected layers and a classifier after the encoding part in the source domain model after unsupervised training to form a classification path of the target domain model; A second adding submodule is used to add 5 deconvolution layers after the decoding part in the source domain model after unsupervised training to form a reconstruction path of the target domain model; The forming submodule is used to form the target domain model by using the classification path and the reconstruction path of the target domain model.

10. The limited label data SAR target classification system according to claim 9, characterized in that: The fine-tuning in the transfer learning module performs the following steps: The decoding part in the reconstruction path is frozen to fine-tune the encoding part in the classification path.

Images