Few-shot SAR automatic target recognition based on Conv-BiLSTM prototypical network

Abstract

In recent studies, synthetic aperture radar (SAR) automatic target recognition (ATR) algorithms have achieved high recognition accuracy in the moving and stationary target acquisition and recognition (MSTAR) data set. However, these algorithms usually require hundreds or more training samples of each target type. In order to extract azimuth-insensitive features in a SAR ATR task with only a few training samples, a convolutional bidirectional long short-term memory (Conv-BiLSTM) network is designed as an embedding network to map the SAR images into a new feature space where the classification problem becomes easier. Based on the embedding network, a novel few-shot SAR ATR framework called Conv-BiLSTM Prototypical Network (CBLPN) is proposed. Experimental results on the MSTAR benchmark data set have illustrated that the proposed method performs well in SAR image classification with only a few training samples.

Introduction

Thanks to its all-day, all-weather, high-resolution and long operating distance capability, synthetic aperture radar (SAR) has been widely applied in battlefield reconnaissance, terrain mapping, geological exploration, and marine observation. Unlike optical imaging, single-polarized SAR image represents the intensity of target scattering by gray scales and usually has blurred edges and strong anisotropy due to background clutter and limited resolution. All these factors will increase the difficulty in effective feature extraction and target recognition.

With recent advances of deep learning theory, deep neural networks have been widely used in many fields [1], [2], [3]. In recent studies, SAR target recognition techniques based on deep learning also have achieved great success with superior performance to traditional ones [4]. However, they usually require a large number of training samples of each target type, which is difficult to meet in real-word situations due to large cost or mission limit. In this scenario, the problem of few-shot SAR ATR arises and invalidates the available algorithms. Therefore, it is quite necessary to take in-depth study on effective few-shot SAR ATR methods.

For SAR target recognition based on single image, there are three mainstream methods, i.e., template matching [5], target modeling [6], and machine learning [7]. Such methods require the designing of a dedicated template, a target model, or a classifier in advance. Nevertheless, the heavy reliance on hand-designed features usually results in high complexity and poor generalization performance.

Recently, deep learning has found wide applications in SAR ATR because of its strong ability in automatic feature extraction, and typical methods mainly include convolutional neural network (CNN) [4], auto-encoders [8], recurrent neural networks [9], and generative adversarial network [10], etc. Among them, CNN and its improved versions have achieved state-of-the-art recognition results in the MSTAR benchmark data set [11]. Their typical structures consist of a feature extractor and a classifier. Specifically, the feature extractor is formed by stacking convolutional layers and pooling layers, and is applied to extract hierarchical features from original data. Initially, a traditional CNN structure which consists of convolutional layers, pooling layers, and the softmax classifier was proposed [12], [13]. Later, its improved versions were developed. For instance, S. Chen et al. designed the A-convent, where the number of unknown parameters was reduced greatly by removing the fully-connected layers [4]. S. Wagner replaced the softmax classifier by the SVM classifier and achieved high recognition accuracy [14]. R. H. Shang et al. added an information recorder to CNN to remember and store the spatial features of samples, and then utilized spatial similarity information of the recorded features to predict the unknown sample labels [15]. J. Wang et al. applied a despeckling sub-network to suppress speckle noise before classification [16]. J. Pei et al. proposed a multi-channel CNN structure, which utilized SAR images with different viewing angels to improve the recognition accuracy [17].

Generally, the available SAR ATR algorithms based on deep learning require a large number of training samples to obtain satisfying generalization performance and alleviate over-fitting. For insufficient training samples, the available techniques may: a) augment the training set by image rotating, shifting and distorting [4], [12]; b) pre-train the network with another data set by transfer learning [13]; or c) design special network structures to reduce the amount of training data required, e.g., to replace the convolutional layers by convolutional highway units [18]. Nevertheless, the above methods still need at least hundreds of training samples for each target type, and their recognition performance will degrade heavily if there are only a few training samples in some classes. Ref. [18] demonstrates that the recognition accuracy of deep learning methods and traditional machine learning methods falls below 40% when the training set only includes dozens of samples for each class (about 10% of the MSTAR SOC training set).

Recently, few-shot learning (FSL) is proposed to tackle the ATR problem where only a few samples of some target types are available [19]. Generally, a few-shot learning task includes three data sets, i.e. the test set, the support set, and the training set. The test set contains the target samples that need to be recognized; the support set contains a few labeled samples that belong to the same classes as the test set; while the training set contains other target classes different from those in the support/test set. By exploiting prior knowledge in the training set, FSL could rapidly generalize to new recognition tasks with limited samples in the support set, mimicking the human ability to acquire knowledge from few examples through generalization. Prototypical Network (PN) [20] is a classical FSL method and has been successfully applied to dermatological disease diagnosis [21] and hyperspectral image classification [22]. This method consists of two main stages: 1) transforming each sample into an embedding vector by a single-channel CNN; and 2) performing classification with the embedding vectors according to the Euclidean distance. Specifically, the unknown parameters in PN are learned by an episode-based method [20]. However, such method cannot be directly applied to SAR ATR. On the one hand, the episode-based network training method requires a large training set with thousands of samples, which is hard to collect in the SAR ATR task. On the other hand, the embedding vectors learned though the single-channel CNN are sensitive to azimuth variation and lack robustness.

To tackle the above-mentioned problems, a novel few-shot SAR ATR method is proposed. The contributions of this paper can be summarized as follows.

a) A novel few-shot learning method for SAR ATR, namely CBLPN, is proposed, which maps each SAR sample to an embedding vector and then performs SAR ATR in the embedding space. Compared with traditional SAR ATR methods, CBLPN obtains close recognition accuracy while requires much less labeled samples.

b) To reduce the influence of azimuth variation on SAR ATR and extract azimuth-robust features from SAR samples, a convolutional bidirectional long short-term memory (Conv-BiLSTM) network is designed to replace the common CNN structure as the feature extractor. Experiments on the MSTAR dataset show that Conv-BiLSTM is less sensitive to azimuth variation of SAR images than CNN and improves the robustness of SAR ATR effectively.

c) A random-episode weights update method is proposed to train the parameters in CBLPN. By randomly sampling in the training set to mimic the real SAR ATR task, the scarcity of labeled samples is alleviated and the parameters in CNLPN can be learned effectively.

The remainder of this paper is organized as follows: Section 2 gives a brief introduction to the research background, including recurrent neural network (RNN) and few-shot learning. Section 3 introduces the structure of Conv-BiLSTM network for SAR feature extraction. Section 4 introduces the framework of CBLPN and explains its components in detail. Section 5 describes the training of CBLPN. Section 6 shows the experimental results with discussions, and finally Section 7 concludes the paper. A summary of the main abbreviations in the paper is listed with their expanded form in Table 1.