Thanks to the medium-to-high resolution and wide coverage imaging ability, satellite synthetic aperture radar (SAR) systems are momentously developed for intelligent maritime surveillance, especially for ship classification in SAR images. Previous researchers mostly utilized the geometric, radiometric, and structural features combined with the traditional machine-learning (ML) methods to conduct ship classification in high-resolution (HR) SAR images. However, the handcrafted features showed weak representation for the medium-resolution (MR) SAR images, and the normal ML methods are less powerful to deal with the intraclass diversity and interclass similarity in the MR SAR ship images. To address these issues, this article extends the dense convolutional networks to the MR SAR ship classification for better deep features extraction and proposes a multitask learning framework within which the softmax log-loss and the triplet loss are jointly minimized for more effective ship classification in MR SAR images. The triplet loss is computed by imposing a similarity constraint on the deep embeddings via the deep metric learning (DML) scheme such that the deep representations of the same class are pulled much closer to each other and those from different classes are pushed as farther apart as possible. However, the triplet loss in the usual DML utilizes the triplets mined in a training batch independently, which ignores the contextual information. So, we propose to impose a Fisher discrimination regularization term on the deep embeddings to explore the global information of learned embeddings, and hence the designed task-specific networks will be trained to have more robust and better recognition performance. Extensive experiments on the MR SAR ship data set collected from the Sentinel-1 SAR images demonstrate that the proposed method can achieve superior performance for the 3- and 5-class recognition tasks with regard to the comparing CNN models.

In this letter, we propose a novel automatic target recognition (ATR) method based on dictionary learning and joint dynamic sparse representation (DL-JDSR) for synthetic aperture radar (SAR) images. First, in the feature extraction step, we extract two kinds of features, i.e., the image domain amplitude feature and the scale-invariant feature transform (SIFT) feature, of which the image domain amplitude feature describes intensity information and the SIFT feature describes gradient information. These two features will be jointly utilized to combine the two kinds of information for SAR ATR. Second, we introduce the dictionary-learning method, the label-consistent K-singular value decomposition, into the training step to learn dictionaries for the two features rather than directly using all training samples as the fixed dictionaries in the traditional sparse representation method. The learned dictionaries have smaller sizes and are more distinctive among different classes, which can speed up our recognition and improve the accuracies. Third, the JDSR algorithm used in the test step employs a more flexible atom selection method, which enables the two features from an image data to share the similar but not exactly the same sparse mode. Experiments on the moving and stationary target acquisition and recognition data set show that the proposed method is an effective way to recognize SAR images.

Synthetic aperture radar (SAR) images are especially susceptible to the target aspect angles. For the SAR target recognition, the lack of training data at different aspect angles inevitably deteriorates the performance. To solve the problem, this letter introduces an angular rotation generative network (ARGN). It is actually an attribute-guided transfer learning method, and the shared attribute between the source and target domains is the target aspect angle. The aspect angle of the data for each type in the source domain covers in the range of 0°-360°, while the information of the aspect angle in the target domain is not complete. Assume that there is a mapping in the feature space between the two images of the same target under different aspect angles, and the mapping relation is learned from sufficient data in the source domain. Then, the mapping can also be applied in the target domain according to the idea of transfer learning. The learned knowledge contained in the feature space helps to improve the target recognition performance in the target domain. Experimental results on the moving and stationary target acquisition and recognition (MSTAR) benchmark data set illustrate that our framework is efficient. For the three-class recognition of the MSTAR data set, the recognition rate is about 87% even when only 23 samples of each class are utilized as the training data. For a given target data, the generation results with counterclockwise rotation of 1°-90° for the aspect angle are performed. The qualitative and quantitative comparisons between the generated images and the real data are also displayed.

Target discrimination has been one of the hottest issues in the interpretation of synthetic aperture radar (SAR) images. However, the presence of speckle noise and the absence of robust features make SAR discrimination difficult to deal with. Recently, convolutional neural network has obtained state-of-the-art results in pattern recognition. In this letter, we propose a target discrimination framework that jointly uses intensity and edge information of SAR images. This framework contains three parts, namely, feature extraction block, feature fusion block, and final classification block. In addition, a novel feature fusion method that can preserve the spatial relationship of different features is introduced. Experimental results on the miniSAR data demonstrate the effectiveness of our method.

To detect ships robustly and automatically in monitoring the marine areas, polarimetric synthetic aperture radar imagery is more and more important. In this letter, three superpixel-level dissimilarity measures are developed to enhance the contrast between ship targets and sea clutter, which are then used to construct an automatic detection algorithm. In the proposed method, multiscale superpixels are first generated. Second, the measurements between a certain superpixel and surrounding ones are calculated. The dissimilarity measures are then transformed from the superpixel level to the pixel level. Third, kernel Fisher discriminant analysis is utilized to improve the separability between ship targets and clutter. Finally, linear support vector machine classifier is utilized to complete the detection automatically. Experiments on the synthetic and real data demonstrate that the proposed method is effective for ship detection with only few false alarms existing, especially under the low signal-to-clutter ratio.

Few-shot synthetic aperture radar (SAR) target classification has received more and more attention in recent years, where most of the existing methods have applied off-the-shelf networks designed for natural images to SAR images, ignoring the special characteristics of SAR data. Therefore, in this article, we propose an attribute-guided multi-scale prototypical network (AG-MsPN) combined with subband decomposition for few-shot SAR target classification, aiming to learn more discriminative features from a few labeled data. Since the SAR images are essentially complex-valued images containing both amplitude and phase information, we implement the subband decomposition of complex-valued SAR images to explore the backscattering variations of targets, thus obtaining more complete descriptions of targets. Then, considering the complementary features extracted by different convolutional layers, based on the prototypical network, a multi-scale prototypical network (MsPN) is proposed to fuse the features of different layers to enhance the discrimination of feature representations, thus relieving the problem of high intra-class diversity and inter-class similarity for the images of SAR targets. Besides, we devise the prior binary attributes of SAR targets and add an extra attribute classification module (ACM) into the MsPN to map the images into the attribute space for classification. During the training phase, the proposed MsPN and the ACM are jointly utilized to realize the target classification in both the feature space and the attribute space, and meanwhile, the model parameters are optimized by the joint loss. Thus, the classification performance of the MsPN is further enhanced under the joint supervision of class label information from a few labeled data and the target attribute information from the prior knowledge. Therefore, we name the proposed method the AG-MsPN. We demonstrate the effectiveness of our proposed AG-MsPN on the Moving and Stationary Target Acquisition and Recognition benchmark dataset, and our method surpasses many other existing methods in the few-shot cases.