When Deep Learning Meets Multi-Task Learning in SAR ATR: Simultaneous Target Recognition and Segmentation

Automatic target recognition (ATR) in synthetic aperture radar (SAR) images has been widely used in civilian and military fields. Traditional model-based methods and template matching methods do not work well under extended operating conditions (EOCs), such as depression angle variant, configuration variant, and noise corruption. To improve the recognition performance, methods based on convolutional neural networks (CNN) have been introduced to solve such problems and have shown outstanding performance. However, most of these methods rely on continuously increasing the width and depth of networks. This adds a large number of parameters and computational overhead, which is not conducive to deployment on edge devices. To solve these problems, a novel lightweight fully convolutional neural network based on Channel-Attention mechanism, Channel-Shuffle mechanism, and Inverted-Residual block, namely the ASIR-Net, is proposed in this paper. Specifically, we deploy Inverted-Residual blocks to extract features in high-dimensional space with fewer parameters and design a Channel-Attention mechanism to distribute different weights to different channels. Then, in order to increase the exchange of information between channels, we introduce the Channel-Shuffle mechanism into the Inverted-Residual block. Finally, to alleviate the matter of the scarcity of SAR images and strengthen the generalization performance of the network, four approaches of data augmentation are proposed. The effect and generalization performance of the proposed ASIR-Net have been proved by a lot of experiments under both SOC and EOCs on the MSTAR dataset. The experimental results indicate that ASIR-Net achieves higher recognition accuracy rates under both SOC and EOCs, which is better than the existing excellent ATR methods.

Section: Recognition Results Under Eocmentioning

confidence: 99%

A Lightweight Fully Convolutional Neural Network for SAR Automatic Target Recognition

Zhou

et al. 2021

“…This SAR image dataset is acquired leveraging the X-band HH polarization "STARLOS" spotlight SAR platform with the resolution of 0.3 m × 0.3 m. As the significant dataset for SAR-ATR performance evaluation, it contains abundant SAR images of vehicle targets and ground clutter. There are ten categories of vehicle targets in the dataset, such as BTR70, BTR60, BRDM2 and BMP2 (armored personnel carrier); 2S1 (rocket launcher); D7 (bulldozer); ZIL131 (truck); T62 and T72(tank); ZSU234 (air defense unit) [33], which are indexed by category labels 1, 2, ..., 10, respectively. These SAR images in each category cover all target-aspect angles in the range of [0 • , 360 • ] with a relative flat grass or exposed soil background.…”

Section: Methodsmentioning

confidence: 99%

Adversarial Attack for SAR Target Recognition Based on UNet-Generative Adversarial Network

Zhang

2021

Some recent articles have revealed that synthetic aperture radar automatic target recognition (SAR-ATR) models based on deep learning are vulnerable to the attacks of adversarial examples and cause security problems. The adversarial attack can make a deep convolutional neural network (CNN)-based SAR-ATR system output the intended wrong label predictions by adding small adversarial perturbations to the SAR images. The existing optimization-based adversarial attack methods generate adversarial examples by minimizing the mean-squared reconstruction error, causing smooth target edge and blurry weak scattering centers in SAR images. In this paper, we build a UNet-generative adversarial network (GAN) to refine the generation of the SAR-ATR models’ adversarial examples. The UNet learns the separable features of the targets and generates the adversarial examples of SAR images. The GAN makes the generated adversarial examples approximate to real SAR images (with sharp target edge and explicit weak scattering centers) and improves the generation efficiency. We carry out abundant experiments using the proposed adversarial attack algorithm to fool the SAR-ATR models based on several advanced CNNs, which are trained on the measured SAR images of the ground vehicle targets. The quantitative and qualitative results demonstrate the high-quality adversarial example generation and excellent attack effectiveness and efficiency improvement.

“…Traditional end to end SAR image unknown target recognition methods [3,6,26,29,31] bear problems such as target feature closure, lack of effective processing, and operation for target features. When faced with unknown targets that are not involved in training, these traditional methods cannot effectively identify the unknown targets.…”

Section: Fea-da Overall Frameworkmentioning

confidence: 99%

“…Most conventional SAR-ATR methods achieve the correct recognition through SAR image target detection, feature extraction and matching the target template library. In recent years, data driven target recognition methods based on deep learning (DL) have been widely used owing to their powerful learning and fitting ability [5][6][7][8]. Chen et al investigated and used the DL network for the recognition of SAR target images, which significantly improved the target recognition performance [9].…”

Section: Introductionmentioning

confidence: 99%

Unknown SAR Target Identification Method Based on Feature Extraction Network and KLD–RPA Joint Discrimination

Zeng

Sun

Xu³

et al. 2021

Recently, deep learning (DL) has been successfully applied in automatic target recognition (ATR) tasks of synthetic aperture radar (SAR) images. However, limited by the lack of SAR image target datasets and the high cost of labeling, these existing DL based approaches can only accurately recognize the target in the training dataset. Therefore, high precision identification of unknown SAR targets in practical applications is one of the important capabilities that the SAR–ATR system should equip. To this end, we propose a novel DL based identification method for unknown SAR targets with joint discrimination. First of all, the feature extraction network (FEN) trained on a limited dataset is used to extract the SAR target features, and then the unknown targets are roughly identified from the known targets by computing the Kullback–Leibler divergence (KLD) of the target feature vectors. For the targets that cannot be distinguished by KLD, their feature vectors perform t-distributed stochastic neighbor embedding (t-SNE) dimensionality reduction processing to calculate the relative position angle (RPA). Finally, the known and unknown targets are finely identified based on RPA. Experimental results conducted on the MSTAR dataset demonstrate that the proposed method can achieve higher identification accuracy of unknown SAR targets than existing methods while maintaining high recognition accuracy of known targets.