SAR Target Classification Method and System for Limited Label Data Based on Transfer Learning of Deep Convolutional Neural Network

Abstract

The present invention discloses a SAR target classification method and system based on transfer learning of a deep convolutional neural network, belonging to the technical field of synthetic aperture radar recognition. The method includes: constructing a source domain model by using a stacked convolutional autoencoder; processing SAR scene images of different unlabeled regions to obtain a source domain data set; using the source domain data set to perform unsupervised training on the source domain model; constructing a target domain model by using the unsupervised trained source domain model; using the augmented labeled target domain data set to perform transfer learning on the target domain model to fine-tune the target domain model; and using the fine-tuned target domain model to perform SAR target classification and recognition. The present invention solves the problem of the small scale of the SAR target data set and improves the accuracy and robustness of SAR target classification.

Description

Technical Field

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

[0002] Synthetic Aperture Radar (SAR) Automatic Target Recognition (SAR-ATR) has important application values in military and civilian fields. Traditional SAR target classification methods mainly rely on manual feature extraction and classifier design, which are difficult to adapt to complex scenarios 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 the SAR target dataset is small, making it 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 based on transfer learning of deep convolutional neural network for limited-label data, which solves the problem of small scale of the SAR target dataset and improves the accuracy and robustness of SAR target classification.

[0004] Another purpose of the present invention is to provide a SAR target classification system based on transfer learning of deep convolutional neural network for limited-label data.

[0005] To achieve the above purpose, the present invention is implemented by adopting the following technical solutions:

[0006] A SAR target classification method based on transfer learning of deep convolutional neural network for limited-label data, the SAR target classification method for limited-label data includes:

[0007] Step S1, constructing a source domain model by using a stacked convolutional autoencoder;

[0008] Step S2, processing SAR scene images of different unlabeled regions to obtain a source domain dataset;

[0009] Step S3, performing unsupervised training on the source domain model by using the source domain dataset;

[0010] Step S4, constructing a target domain model by using the source domain model after unsupervised training;

[0011] Step S5, performing transfer learning on the target domain model by using the augmented labeled target domain dataset to fine-tune the target domain model;

[0012] Step S6, performing SAR target classification and recognition by using the fine-tuned target domain model.

[0013] Further, in the step S1, the source domain model includes an encoding part and a decoding part of a stacked convolutional autoencoder;

[0014] The encoding part includes a plurality of convolutional layers with the number of feature map channels increasing sequentially; after each convolutional layer, an activation function and a max pooling layer are connected in sequence;

[0015] The decoding part includes deconvolutional layers corresponding to the respective convolutional layers; each deconvolutional layer is used for upsampling; an activation function is connected after each deconvolutional layer.

[0016] Further, in the step S2, the regions include urban regions, forest regions, mountain regions, and cultivated land regions.

[0017] Further, in the step S4, the construction process of the target domain model is as follows:

[0018] Step S41, add two fully connected layers and a classifier after the encoding part in the source domain model after unsupervised training to form the classification path of the target domain model;

[0019] Step S42, add 5 deconvolutional layers after the decoding part in the source domain model after unsupervised training to form the reconstruction path of the target domain model;

[0020] Step S43, use the classification path and the reconstruction path of the target domain model to form the target domain model.

[0021] Further, in the step S5, the specific process of the fine-tuning is as follows:

[0022] Freeze the decoding part in the reconstruction path and fine-tune the encoding part in the classification path.

[0023] To achieve the second above-mentioned object, the present invention is implemented by adopting the following technical solution:

[0024] A limited-label data SAR target classification system based on deep convolutional neural network transfer learning, the limited-label data SAR target classification system includes:

[0025] A first construction module for constructing a source domain model by using a stacked convolutional autoencoder;

[0026] A processing module for processing SAR scene images of different unlabeled regions to obtain a source domain data set;

[0027] An unsupervised training module for performing unsupervised training on the source domain model by using the source domain data set;

[0028] A second construction module for constructing a target domain model by using the source domain model after unsupervised training;

[0029] A transfer learning module for performing transfer learning on the target domain model by using the augmented labeled target domain data set to fine-tune the target domain model;

[0030] A classification and recognition module for performing SAR target classification and recognition by using the fine-tuned target domain model.

[0031] Furthermore, the source domain model includes an encoding part and a decoding part of a stacked convolutional autoencoder;

[0032] The encoding part includes a plurality of convolutional layers with the number of feature map channels increasing in sequence; after each convolutional layer, an activation function and a maximum pooling layer are connected in sequence;

[0033] The decoding part includes deconvolutional layers corresponding to the respective convolutional layers; each deconvolutional layer is used for upsampling; an activation function is connected after each deconvolutional layer.

[0034] Furthermore, the regions include urban regions, forest regions, mountain regions, and cultivated land regions.

[0035] Furthermore, the second construction module includes:

[0036] A first addition sub-module for 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;

[0037] A second addition sub-module for adding 5 deconvolutional layers after the decoding part in the source domain model after unsupervised training to form a reconstruction path of the target domain model;

[0038] A formation sub-module for forming the target domain model by using the classification path and the reconstruction path of the target domain model.

[0039] Furthermore, the fine-tuning in the transfer learning module is performed as follows:

[0040] Freeze the decoding part in the reconstruction path to fine-tune the encoding part in the classification path.

[0041] In summary, the technical solution of the present invention has the following technical effects:

[0042] The present invention uses SAR scene images of different unlabeled regions (i.e., source domain datasets) to perform unsupervised training on a source domain model constructed by a stacked convolutional autoencoder to learn the hierarchical feature representation of SAR scene images; and uses the source domain model after unsupervised training to construct a target domain model; uses the augmented labeled target domain dataset 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, solving the problem of the small scale of the SAR target dataset and the difficulty in training a large CNN model, and improving the accuracy and robustness of SAR target classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for use in the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0044] Figure 1 It is a schematic flowchart of a SAR target classification method based on transfer learning of a deep convolutional neural network according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0046] This embodiment provides a SAR target classification method based on transfer learning of a deep convolutional neural network. Refer to Figure 1 , the SAR target classification method with limited labeled data includes:

[0047] Step S1: Use a stacked convolutional autoencoder to construct a source domain model.

[0048] 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 a maximum pooling layer with a ReLU activation function and , and adopts a pyramid structure, with the number of feature map channels increasing sequentially. After each convolutional layer, a ReLU activation function and a The max pooling layer. The decoding part is symmetric to the encoding part in structure. It uses a transposed convolutional layer for upsampling and connects a ReLU activation function. During the training process, the mean squared error (MSE) is used as the loss function, and the mini-batch stochastic gradient descent (SGD) algorithm is used for optimization.

[0049] Based on the above concept, the source domain model of this embodiment includes the encoding part and the decoding part of the stacked convolutional autoencoder;

[0050] The encoding part includes multiple convolutional layers with the number of feature map channels increasing sequentially; After each convolutional layer, an activation function and a max pooling layer are connected in sequence;

[0051] The decoding part includes transposed convolutional layers corresponding to each convolutional layer; Each transposed convolutional layer is used for upsampling; An activation function is connected after each transposed convolutional layer.

[0052] Step S2: Process the SAR scene images of different unlabeled regions to obtain the source domain dataset.

[0053] Construct a source domain dataset from a large number of unlabeled SAR scene images collected by the German Earth observation satellite TerraSAR-X, covering regions with various landscapes such as cities, forests, mountains, and cultivated lands (that is, the regions include urban regions, forest regions, mountain regions, and cultivated land regions.), and crop them into pixels, and perform preprocessing operations such as normalization to obtain the source domain dataset.

[0054] Step S3: Use the source domain dataset to perform unsupervised training on the source domain model.

[0055] In this embodiment, a stacked convolutional autoencoder (SCAE) is used to perform unsupervised training on the source domain dataset to learn the hierarchical feature representation of SAR scene images, and obtain the source domain model after unsupervised training.

[0056] During the training process, the mean squared error (MSE) is used as the loss function, and the mini-batch stochastic gradient descent (SGD) algorithm is used for optimization.

[0057] Step S4: Use the source domain model after unsupervised training to construct the target domain model.

[0058] 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 5 transposed convolutional layers after the encoding part of the source domain model for reconstructing the SAR target image. The construction process of the target domain model is as follows:

[0059] Step S41: Add two fully connected layers and a classifier after the encoding part in the source domain model after unsupervised training to form the classification path of the target domain model, and achieve SAR target classification;

[0060] Step S42: Add 5 deconvolution layers after the decoding part in the source domain model after unsupervised training to form the reconstruction path of the target domain model, and achieve SAR target image reconstruction;

[0061] Step S43: Use the classification path and reconstruction path of the target domain model to form the target domain model.

[0062] Step S5: Use the augmented labeled target domain dataset to perform transfer learning on the target domain model to fine-tune the target domain model.

[0063] Select 10 classes of vehicles from the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset: images of T72, BTR70, BMP2, 2S1, BRDM2, BTR60, D7, T62, ZIL131, and ZSU23 to construct the target domain dataset. The images contain data at two angles of 15° and 17°. Perform preprocessing operations such as cropping and normalization on the images, and perform data augmentation, such as translation and mirroring, to obtain the augmented labeled target domain dataset.

[0064] Transfer the feature representations learned in the source domain model to the target domain model and perform fine-tuning to adapt to the SAR target classification task. During the transfer learning process, freeze the encoding part of the source domain model, and only fine-tune the decoding parts of the classification path and reconstruction path of the target domain model. The specific process of fine-tuning is as follows:

[0065] Freeze the decoding part in the reconstruction path and fine-tune the encoding part in the classification path.

[0066] Step S6: Use the fine-tuned target domain model to perform SAR target classification and recognition.

[0067] Use the target domain model after transfer learning to classify and recognize SAR targets.

[0068] The classification results are evaluated using a confusion matrix, and metrics such as accuracy, recall, F1 value, and mean average precision (mAP) are calculated. To verify the effectiveness of the method of the present invention, it is compared with the following methods:

[0069] 1. SVM: Use a support vector machine for SAR target classification and train using manually extracted features.

[0070] 2. MSRC: Use the single-gene signal sparse representation method for SAR target classification.

[0071] 3. TJSR: Use the three-task joint sparse representation method for SAR target classification.

[0072] 4. SDDLSR: Use the supervised discriminant dictionary learning and sparse representation method for SAR target classification.

[0073] 5. JDSR: Use the joint dynamic sparse representation method for SAR target classification.

[0074] 6. ConvNet: Use the fully convolutional network for SAR target classification and train from scratch.

[0075] 7. DCHUN: Use the deep convolutional neural network for SAR target classification and fine-tune using a pre-trained model.

[0076] The experimental results show that the classification accuracy, recall rate, F1 value, mAP and other indicators of this embodiment on the MSTAR dataset are better than the above comparison methods. Especially when the dataset size is small, the performance improvement of the method of the present invention is more obvious.

[0077] This embodiment uses SAR scene images of different unlabeled regions (i.e., the source domain dataset) to perform unsupervised training on the source domain model constructed by using a stacked convolutional autoencoder to learn the hierarchical feature representation of SAR scene images; and uses the source domain model after unsupervised training to construct a target domain model; uses the augmented target domain dataset with labels 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, solving the problems of small SAR target dataset size and difficulty in training large CNN models, and improving the accuracy and robustness of SAR target classification.

[0078] The above embodiment can be implemented by adopting the technical solution of the following embodiment:

[0079] Another embodiment provides a limited-label data SAR target classification system based on deep convolutional neural network transfer learning. The limited-label data SAR target classification system includes:

[0080] The first construction module is used to construct a source domain model by using a stacked convolutional autoencoder;

[0081] The processing module is used to process SAR scene images of different unlabeled regions to obtain a source domain dataset;

[0082] The unsupervised training module is used to perform unsupervised training on the source domain model by using the source domain dataset;

[0083] A second construction module, configured to construct a target domain model by using the source domain model after unsupervised training;

[0084] A transfer learning module, configured to perform transfer learning on the target domain model by using the augmented labeled target domain data set, so as to fine-tune the target domain model;

[0085] A classification and recognition module, configured to perform SAR target classification and recognition by using the fine-tuned target domain model.

[0086] Furthermore, the source domain model includes an encoding part and a decoding part of a stacked convolutional autoencoder;

[0087] The encoding part includes a plurality of convolutional layers with the number of feature map channels increasing sequentially; after each convolutional layer, an activation function and a maximum pooling layer are connected in sequence;

[0088] The decoding part includes deconvolutional layers corresponding to the respective convolutional layers; each deconvolutional layer is used for upsampling; an activation function is connected after each deconvolutional layer.

[0089] Furthermore, the regions include urban regions, forest regions, mountain regions, and cultivated land regions.

[0090] Furthermore, the second construction module includes:

[0091] A first addition sub-module, configured to add two fully-connected layers and a classifier after the encoding part in the source domain model after unsupervised training, so as to form a classification path of the target domain model;

[0092] A second addition sub-module, configured to add 5 deconvolutional layers after the decoding part in the source domain model after unsupervised training, so as to form a reconstruction path of the target domain model;

[0093] A formation sub-module, configured to form the target domain model by using the classification path and the reconstruction path of the target domain model.

[0094] Furthermore, the fine-tuning in the transfer learning module is performed as follows:

[0095] Freeze the decoding part in the reconstruction path to fine-tune the encoding part in the classification path;

[0096] The principles, formulas, and their parameter definitions involved in the above embodiments are all applicable and will not be elaborated here one by one.

[0097] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification. The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A SAR target classification method for limited-label data based on transfer learning of deep convolutional neural network, characterized in that, The limited-label data SAR target classification method includes: Step S1: Use a stacked convolutional autoencoder to construct a source domain model; Step S2: Process SAR scene images of different unlabeled regions to obtain a source domain dataset; Step S3: Use the source domain dataset to perform unsupervised training on the source domain model to learn the hierarchical feature representation of SAR scene images; Step S4: Use the source domain model after unsupervised training to construct a target domain model; In Step S4, the construction process of the target domain model is as follows: Step S41: Add two fully connected layers and a classifier after the encoding part of the source domain model after unsupervised training to form the classification path of the target domain model; Step S42: Add 5 deconvolutional layers after the decoding part of the source domain model after unsupervised training to form the reconstruction path of the target domain model; Step S43: Use the classification path and reconstruction path of the target domain model to form the target domain model; Step S5: Transfer the hierarchical feature representation of SAR scene images learned in the source domain model to the target domain model, and use the augmented labeled target domain dataset 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 tag data SAR target classification method according to claim 1, characterized in that, In Step S1, the source domain model includes the encoding part and the decoding part of the stacked convolutional autoencoder; The encoding part includes a plurality of convolutional layers with the number of channels of the feature maps increasing sequentially; after each convolutional layer, an activation function and a max-pooling layer are connected in sequence; The decoding part includes deconvolutional layers corresponding to each convolutional layer; each deconvolutional layer is used for upsampling; an activation function is connected after each deconvolutional layer.

3. The limited tag data SAR target classification method according to claim 2, wherein In Step S2, the regions include urban regions, forest regions, mountain regions, and cultivated land regions.

4. The method for classifying SAR targets with limited tag data according to claim 3, characterized in that, In Step S5, the specific process of the fine-tuning is as follows: Freeze the decoding part in the reconstruction path and fine-tune the encoding part in the classification path.

5. A SAR target classification system for limited-label data based on transfer learning of deep convolutional neural network, characterized in that, The limited-label data SAR target classification system includes: The first construction module is used to construct a source domain model using a stacked convolutional autoencoder; The processing module is used to process SAR scene images of different unlabeled regions to obtain a source domain dataset; The unsupervised training module is used to perform unsupervised training on the source domain model using the source domain dataset to learn the hierarchical feature representation of SAR scene images; The second construction module is used to construct a target domain model using the source domain model after unsupervised training; The second construction module includes: The first addition sub-module is used to add two fully connected layers and a classifier after the encoding part of the source domain model after unsupervised training to form the classification path of the target domain model; The second addition sub-module is used to add 5 deconvolutional layers after the decoding part of the source domain model after unsupervised training to form the reconstruction path of the target domain model; The formation sub-module is used to form the target domain model using the classification path and reconstruction path of the target domain model; The transfer learning module is used to transfer the hierarchical feature representation of SAR scene images learned in the source domain model to the target domain model, and use the augmented 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 by using the fine-tuned target domain model.

6. The limited tag data SAR target classification system according to claim 5, wherein 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 the number of channels of the feature maps increasing sequentially; after each convolutional layer, an activation function and a max pooling layer are connected in sequence; The decoding part includes deconvolutional layers corresponding to each convolutional layer; each deconvolutional layer is used for upsampling; an activation function is connected after each deconvolutional layer.

7. The limited tag data SAR target classification system according to claim 6, wherein The regions include urban regions, forest regions, mountain regions, and cultivated land regions.

8. The limited tag data SAR target classification system according to claim 7, characterized in that The fine-tuning in the transfer learning module is performed as follows: Freeze the decoding part in the reconstruction path to fine-tune the encoding part in the classification path.

Images