In this letter, the single shot multibox detector (SSD), which is a real-time object detection method based on convolutional neural network, is applied to realize target detection for synthetic aperture radar (SAR) images. Since there are no sufficient labeled images for training in SAR target detection, we apply two strategies, data augmentation and transfer learning. For data augmentation, the first approaches to use some image processing methods, i.e., manual-extracting subimages, adding noise, filtering, and flipping, on the original training images to generate some new training images; the second approach is to employ the existing SAR target recognition data set, MSTAR data set, to assist in accomplishing the target detection task. For transfer learning, we first apply subaperture decomposition technique on original SAR images to acquire three-channel subaperture SAR images, and then transfer the three-channel VGGNet model pretrained on the ImageNet data set to the three-channel subaperture SAR images, in order to initialize corresponding parameters of the convolutional layers in the base network in our SSD. The feature extraction network, consisting of the base network and the auxiliary structure, is used to learn multiscale feature maps, and then convolutional predictors are used to acquire the final detection results. The experimental results on the miniSAR real image data set demonstrate that the proposed method can obtain better detection performance than other detection methods.

This article proposes a multiscale convolutional neural network (CNN) based on component analysis (CA-MCNN) for synthetic aperture radar (SAR) automatic target recognition (ATR). The component information of a target is robust to the local variations of the target, which is not made the best of by traditional CNN-based methods. For learning the component information, we use the attributed scattering centers (ASCs) extracted from the target echoes as the components of the target for SAR ATR, which divides the SAR target according to the geometric scattering types of ASCs and can not only make the division results more robust but also accurately characterize the electromagnetic scattering characteristics of the target. Since the global information provided by the whole image is also important for SAR ATR, CA-MCNN combines the global information with component information to learn a more efficient and robust target feature representation. In addition, considering that the feature maps of the shallower layer in CNN focus on local and fine-grained information while the feature maps in the deeper layer focus on global and coarse-grained information, we fuse the multiscale feature maps obtained from different layers to enhance the feature description ability. Extensive experiments conducted on the moving and stationary target acquisition and recognition (MSTAR) data set prove the superior performance of CA-MCNN.

Target detection for synthetic aperture radar (SAR) images has great influence on the successive discrimination based on the target regions. However, as a pixel-based method, the traditional constant false alarm rate (CFAR) detection could not work well for the ship target detection problem of multiple ship targets with different sizes in a SAR image, which is referred to as the multiscale situation. Moreover, it needs to use the clustering method on the pixel-level detection results to obtain the accurate target regions, which may merge two or more different targets into a target region. In this letter, a modified CFAR based on object proposals is proposed. We use the object proposal generator to generate a small set of object proposals with different sizes, and then use the proposal-based CFAR detector, where the extracted object proposals are regarded as the guard windows instead of setting fixed guard window, to detect the true positive object proposals. By introducing the object proposals as the variable guard windows in the CFAR detector, the proposed algorithm could gain good detection performance in the multiscale situation, since the missed detection resulting from the big differences between the sizes of the fixed guard window and ship targets can be avoided. Meanwhile, the proposed method can directly obtain the accurate target regions. The effectiveness of the proposed algorithm is verified using the measured SAR data.

The single shot multibox detector (SSD), a proposal-free method based on convolutional neural network (CNN), has recently been proposed for target detection and has found applications in synthetic aperture radar (SAR) images. Moreover, the saliency information reflected in the saliency map can highlight the target of interest while suppressing clutter, which is beneficial for better scene understanding. Therefore, in this article, we propose a saliency-guided SSD (S-SSD) for target detection in SAR images, in which we effectively integrate the saliency into the SSD network not only to suggest where to focus on but also to improve the representation capability in complex scenes. The proposed S-SSD contains two separated convolutional backbone subnetwork architectures, one with the original SAR image as input to extract features, and the other with the corresponding saliency map obtained from the modified Itti's method as input to acquire refined saliency information under supervision. In addition, the dense connection structure, instead of the plain structure used in original SSD, is applied in the two convolutional backbone architectures to utilize multiscale information with fewer parameters. Then, for integrating saliency information to guide the network to emphasize informative regions, multilevel fusion modules are utilized to merge the two streams into a unified framework, thereby making the whole network end-to-end jointly trained. Finally, the convolutional predictors are used to predict targets. The experimental results on the miniSAR real data demonstrate that the proposed S-SSD can achieve better detection performance than state-of-the-art methods.

In recent years, synthetic aperture radar (SAR) ship detection methods based on convolutional neural networks (CNNs) have attracted wide attention in remote sensing fields. However, these methods require a large number of labeled SAR images to train the network, where labeling for SAR images is more expensive and time-consuming than that for optical images. To address the problem of lacking labeled SAR images, in this article, we proposed an unsupervised domain adaptation (UDA) framework based on progressive transfer for SAR ship detection by transferring knowledge from the optical domain to the SAR domain. Due to the prominent difference between the optical and SAR images, our approach progressively transfers knowledge from three levels: pixel level, feature level, and prediction level. At the pixel level, considering the difference in imaging mechanism, we propose a special data augmentation method for ship targets and build the generator with skip connection based on generative adversarial networks (GANs) to generate transition domain, which can reduce the appearance discrepancy between the optical and SAR images. At the feature level, the detector is trained to learn the domain-invariant features with adversarial alignment. At the prediction level, we further use the predicted pseudo-labels obtained by the feature-aligned detector to learn more discriminative features of the SAR images directly and propose the robust self-training (RST) method to reduce the influence of noisy pseudo-labels on detector training. Specially, RST is formulated as a loss minimization problem for object detection. The experimental results based on the domain adaptation from optical dataset to SAR dataset demonstrate that our approach achieves superior SAR ship detection performance with unlabeled SAR images.

In the remote sensing image processing field, the synthetic aperture radar (SAR) target-detection methods based on convolutional neural networks (CNNs) have gained remarkable performance relying on large-scale labeled data. However, it is hard to obtain many labeled SAR images. Semi-supervised learning is an effective way to address the issue of limited labels on SAR images because it uses unlabeled data. In this paper, we propose an improved faster regions with CNN features (R-CNN) method, with a decoding module and a domain-adaptation module called FDDA, for semi-supervised SAR target detection. In FDDA, the decoding module is adopted to reconstruct all the labeled and unlabeled samples. In this way, a large number of unlabeled SAR images can be utilized to help structure the latent space and learn the representative features of the SAR images, devoting attention to performance promotion. Moreover, the domain-adaptation module is further introduced to utilize the unlabeled SAR images to promote the discriminability of features with the assistance of the abundantly labeled optical remote sensing (ORS) images. Specifically, the transferable features between the ORS images and SAR images are learned to reduce the domain discrepancy via the mean embedding matching, and the knowledge of ORS images is transferred to the SAR images for target detection. Ultimately, the joint optimization of the detection loss, reconstruction, and domain adaptation constraints leads to the promising performance of the FDDA. The experimental results on the measured SAR image datasets and the ORS images dataset indicate that our method achieves superior SAR target detection performance with limited labeled SAR images.