Multilevel Adaptive Knowledge Distillation Network for Incremental SAR Target Recognition

Combination of global and local filters for robust SAR target recognition under various extended operating conditions

DARPA/Air Force Research Laboratory Moving and Stationary Target Acquisition and Recognition (MSTAR) program is developing state-of-the-art model based vision approach to Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR). The model-based approach requires using off-line developed target models in an on-line hypothesize-and-test manner to compare predicted target signatures with image data and output target reports. Central to this model-based ATR is the PEMS (Predict-Extract-Match-Search) subsystem. The Search module is critical to PEMS by providing intelligent control to traverse the hypothesis feature space. A major MSTAR goal is to demonstrate robust ATR for variations in targets including partially hidden targets. This paper will provide an update on the technology being developed under MSTAR and the status of this model based ATR research, specifically concentrating on the Search Module.

In the field of synthetic aperture radar (SAR) automatic target recognition (ATR), deep learning-driven methods perform well when numerous training samples are available. However, in practical SAR application scenarios, the number of available training samples is often scarce due to the limitation of observation conditions, which leads to major challenges for deep learning-driven methods in real-world deployment. Currently, the problem of ATR with only a few training samples has attracted much attention in the SAR field. In this article, we put forward an ATR method called azimuth-guided feature embedding network with dual inference mechanism (AGFEN-DIM) to achieve fewshot SAR target recognition. To be specific, the feature extraction model, i.e. azimuth-guided feature embedding network (AGFEN) is composed of the azimuth embedding module (AEM) and the dynamic feature embedding module (DFEN). Among them, AEM resorts to a set of azimuth-parameterized convolution kernels, aiming to make full use of information relevant to the current radar imaging environment. The proposed DFEN is capable of dynamically extracting the discriminative features of the target according to the input image under the guidance of azimuth-parameterized convolutional kernels. The dual inference mechanism (DIM) effectively integrates the inference results of improved prototype classification and label propagation strategy to boost the reliability and robustness of the recognition model for few-shot SAR target recognition tasks. Extensive experiments on the moving and stationary target acquisition and recognition (MSTAR) dataset demonstrate that the proposed AGFEN-DIM surpasses many state-of-the-art SAR ATR methods in various application scenarios.

In recent years, convolutional neural network (CNN) has been increasingly considered as a promising technology for military and homeland security applications. The fusion of CNN and Support vector machine (SVM), a popular traditional machine learning approach, has received intensive attention in the field of synthetic aperture radar (SAR) automatic target recognition (ATR). This paper, firstly, discusses the effects of some preprocessing and image enhancement methods on the performance of SAR ATR, starting with the pre-trained AlexNet in a transfer-learning based approach. Secondly, the architecture of AlexNet is modified to form a new model suitable for SAR ATR. Finally, we propose a hybrid model associated with the success of the learning feature of our CNN model and the ability of SVM to process high-dimensional dataset effectively. To evaluate the proposed method, experiments are performed on the Moving and Stationary Target Acquisition and Recognition (MSTAR) public database. The comparative results demonstrate that these preprocessing and enhancement methods prior to the deep-learning process are not necessary since the feature representation ability of AlexNet is already powerful. Furthermore, experimental results on the benchmark MSTAR dataset illustrate the effectiveness of the proposed new model. On classification of ten-class targets, the commonly used translation augmentation for training data has been performed. By combining the CNN and SVM, the classification accuracy percentages can be slightly improved for our proposed new model.

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

Synthetic aperture radar (SAR) automatic target recognition is an important application in SAR. How to extract features has restricted the application of SAR technology seriously. In this paper, a new feature extraction method for SAR automatic target recognition based on maximum interclass distance is proposed, which integrates class and neighborhood information. This method can reinforce discriminative power using maximum interclass distance, so it can improve recognition rate effectively.

Attributed scattering center is one of important features of Synthetic Aperture Radar (SAR) image. In this paper, a method for the matching of attributed scattering centers and its application to SAR target recognition is proposed. First, the attributed scattering centers of the test SAR image and template SAR images are extracted on the basis of the attributed scattering model. Second, the Hungarian algorithm is employed to match the two scattering center sets. Based on the one to one correspondence, we design a new similarity measure to evaluate the similarity between the two scattering center sets that will decide the target type of the test image. The similarity measure considers the effects of each individual scattering center, single matching pair, and missing alarms and false alarms; thus, it is more comprehensive. The experiment based on moving and stationary target acquisition and recognition database demonstrates the validity of the proposed method.

Data augmentation is one of the most straightforward methods to alleviate the problem of data scarcity in deep learning-based synthetic aperture radar (SAR) automatic target recognition (ATR). For example, novel view generation based on physical electromagnetic models or deep generative models. Besides the data scale, the quality of generated SAR images also has a significant impact on the model's performance. Therefore, it is crucial to objectively and effectively assess the quality of generated SAR images. In this work, a relative recognition accuracy score (RRS)-based evaluation criterion is proposed. The core idea of the proposed evaluation criterion is to control variables. First, the measured SAR images are gradually replaced by the generated SAR images at a progressive hybrid rate (HR) to obtain hybrid training datasets with varying HR levels. Then, the recognition model (e.g., AconvNet) is trained on both the all-measured dataset and hybrid datasets with varying HR to plot the reference recognition accuracy (RRA) and the hybrid recognition curve (HRC). Finally, the integral difference between RRA and HRC is defined as the RRS of the generated SAR dataset. Specifically, we employ proportional random sampling and multiple sampling averaging to alleviate selection bias. Experiments on the moving and stationary target acquisition and recognition (MSTAR) dataset validate the rationality and effectiveness of the proposed RRS-based evaluation criterion.

This paper proposes an automatic target recognition (ATR) method for synthetic aperture radar (SAR) images based on information-decoupled representation. A typical SAR image of a ground target can be divided into three parts: target region, shadow and background. From the aspect of SAR target recognition, the target region and shadow contain discriminative information. However, they also include some confusing information because of the similarities of different targets. The background mainly contains redundant information, which has little contribution to the target recognition. Because the target segmentation may impair the discriminative information in the target region, the relatively simpler shadow segmentation is performed to separate the shadow region for information decoupling. Then, the information-decoupled representations are generated, i.e., the target image, shadow and original image. The background is retained in the target image, which represents the coupling of target backscattering and background. The original image and generated target image are classified using the sparse representation-based classification (SRC). Then, their classification results are combined by a score-level fusion for target recognition. The shadow image is not used because of its lower discriminability and possible segmentation errors. To evaluate the performance of the proposed method, extensive experiments are conducted on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset under both standard operating condition (SOC) and various extended operating conditions (EOCs). The proposed method can correctly classify 10 classes of targets with the percentage of correct classification (PCC) of 94.88% under SOC. With the PCCs of 93.15% and 75.03% under configuration variance and 45° depression angle, respectively, the superiority of the proposed is demonstrated in comparison with other methods. The robustness of the proposed method to both uniform and nonuniform shadow segmentation errors is validated with the PCCs over 93%. Moreover, with the maximum average precision of 0.9580, the proposed method is more effective than the reference methods on outlier rejection.

In realistic synthetic aperture radar (SAR) scenarios, new categories of targets continue to appear over time. It requires the automatic target recognition (ATR) model should continue to accommodate and recognize new categories of targets while maintaining stability on old categories of targets. Therefore, how to strike a glorious balance between plasticity and stability is a matter of wide concern for incremental SAR ATR. In this letter, we propose a new incremental ATR method called dynamic embedding relation distillation network (DERDN), which strives to alleviate the dilemma between plasticity and stability. Specifically, we first develop a dynamic feature embedding model whose parameters can be adjusted adaptively with the data, which is very conducive to continuous generalization to new category targets. To mitigate catastrophic forgetting of model on old categories, we then propose a relation distillation strategy based on exemplar replay to recall knowledge of old categories, thereby improving the stability of the model. Moreover, a hybrid loss that takes into account both label space and embedding space is presented, which ensures inter-class discriminability while enhancing intra-class compactness in the feature space, so as to sustainably accommodate new categories. Evaluation experiments on two benchmark datasets, i.e., moving and stationary target acquisition and recognition (MSTAR) and the synthetic and measured paired labeled experiment (SAMPLE), illustrate that the proposed method is competitive with many state-of-the-arts. The recognition rate of the proposed method is up to 3% higher than that of the best competitor, and its forgetting rate is about 2.5% lower than that of the best competitor.

Recently, significant advancements have been made in few-shot learning (FSL) methods based on metric learning, which have been widely applied to Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR). These methods typically require experimental samples to exhibit sufficiently small intra-class distances. However, in practical applications, SAR target samples are often acquired under varying depression angles, resulting in mixed images. The inherent feature deviations caused by variations in depression angles lead to increased intra-class distances in the feature distribution, which adversely affects recognition performance. To address this challenge, we propose a novel approach called the Dense Graph Prototype Network (DGP-Net). DGP-Net addresses the feature deviation problem by learning potential features and utilizes feature distribution modeling for classification. Specifically, by leveraging the information propagation mechanism of a densely connected Graph Convolutional Network (GCN), potential features are iteratively learned while retaining previous features, thereby clustering samples of the same class with different elevation angles and eliminating feature deviations. To avoid sampling examples located at the edges of the distribution in FSL tasks, which may lead to inaccurate GCN information propagation, we introduce a ”prototype node” mechanism to enhance the robustness of the model. We conduct experiments using samples from the Moving and Stationary Target Acquisition and Recognition (MSTAR) benchmark dataset with varying elevation angles for the same class. Experimental results on this dataset demonstrate the superior performance of DGP-Net. Specifically, in 3-way and 5-way classification tasks, the recognition accuracy of DGP-Net improved by over 4% compared to state-of-the-art (SOTA) methods. The source code for DGP-Net is available at https://github.com/Zhou-xy99/DGP-Net.

A synthetic aperture radar (SAR) automatic target recognition can effectively improve the utilization efficiency of SAR image data. In order to improve the generalization and accuracy of SAR image target recognition, a convolutional neural network model is proposed to be applied to SAR image target recognition. Firstly, the network structure is designed for the characteristics of SAR image. Secondly, the introduction of dense block network structure improves the generalization performance of the model. Finally, Dropout reduced the computational complexity of the network and improved the generalization performance of the model. Experimental data were obtained from the United States Moving and Stationary Target Acquisition and recognition database. Experimental results of 10 types of target recognition showed that the overall recognition rate of the improved convolutional neural network is 99.18%. The convolutional neural network model proposed in this study improves the accuracy and generalization of the network, and is an effective method for target recognition of SAR images.

The existing deep learning studies on synthetic aperture radar (SAR) automatic target recognition (ATR) mainly focus on the utilization of the amplitude of SAR image via convolutional neural network (CNN), while the electromagnetic scattering information is rarely considered. Given that scattering centers (SCs) can characterize the target’s electromagnetic scattering characteristics and describe the target’s physical structure information, the SC feature should be helpful for SAR ATR. Therefore, we propose a novel SAR ATR method that combines electromagnetic scattering information and graph convolutional network (GCN) effectively and directly. Specifically, we model each extracted SC as a node to convert the SCs into graph data. The constructed graph is learned via GCN to describe the target’s physical structure information, where the features of different GCN layers are fused to avoid the over-smoothing of GCN. Label smoothing is combined with GCN for the first time to alleviate the overfitting caused by the limited training data. To the best of our knowledge, this study is the first to introduce the GCN for effectively utilizing the SCs, proving that the structural characteristics of the SCs of SAR targets are remarkably beneficial for recognition. Extensive experimental results on the measured moving and stationary target acquisition and recognition (MSTAR) dataset show that our method can obtain superior recognition performance compared with the existing methods.

The traditional Synthetic Aperture Radar Automatic Target Recognition (SAR/ATR) algorithm can train a sufficient number of known class samples and classify the samples in the test set. However, if the old model is trained only with the new class samples, the old class samples’ knowledge is easily forgotten by the new model, which is called catastrophic forgetting. The reason is that the model only fits the distribution of current training samples, so training the whole data set is necessary. Due to the limitation of storage resources, it is often not feasible to retain the whole data set. In order to avoid this phenomenon, a small number of old class samples can be kept to train with the new class samples. Therefore, how to select the old class samples becomes the key point. In this paper, the Density Coverage-Based Exemplar Selection (DCBES) is proposed to choose the key samples of the old class. DCBES selects samples based on the metric learning theory and the set covering theory. First, the metric learning theory is used to measure the similarity between samples and to obtain the density range of samples. Then the exemplar selection problem is considered a set covering problem, to select a fixed number of exemplars to achieve the maximum coverage of the class density range. Experiments on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset show that our method is superior to other exemplar selection methods and achieves the best results.

A new approach is developed for synthetic aperture radar (SAR) automatic target recognition based on convolutional neural network (CNN). In order to improve the low efficiency caused by fixed learning rate in CNN, the AdaDelta method is introduced to adjust the learning rate adaptively. The experimental results on Moving and Stationary Target Acquisition and Recognition (MSTAR) data sets show that convergence is significantly faster in proposed method than CNN with fixed learning rate.