Recently, the deep learning models have achieved great success in the recognition of inverse synthetic aperture radar (ISAR) images. However, most of the deep learning models fail to obtain satisfactory results under the condition of small samples due to the contradiction between the large parameter space of the deep learning models and the insufficient labeled samples of space target imaging by ISAR. In this article, a method of meta-learner-based stacking network (MSN) is proposed, which can realize the high-precision classification of space target by ISAR images under the condition of small sample. Innovatively, a rotation-invariant attention mechanism (RAM) module is added into Resnet50 network to magnify the difference of embedded features of target and background. Complementarily, the deep relationship between the features of fine-grained ISAR image is extracted by using graph convolutional network and relation network. Finally, an innovative adaptive weighted XGBoost algorithm is used to integrate the prediction results of the base learners. The main contributions of this article include proposing a RAM module and using an innovative adaptive weighted XGBoost algorithm to realize ensemble learning. The experiment results show that the RAM module effectively concentrates the network's attention on the recognized target, and the recognition rate of MSN is about 5% higher than that of a single base learner under different data volume conditions, which proves that MSN achieves competitive accuracy against other state-of-the-art approaches.

In recent years, research on the recognition for inverse synthetic aperture radar (ISAR) images continues to deepen, while most methods only use the amplitude information of the ISAR image data. Besides, high-order terms in the complex-valued (CV) received signals for maneuvering space targets will cause defocusing on the ISAR images, which affects the accuracy of the recognition. For a steadily rotating maneuvering target, its high-order phase information between frames is relevant, and this information can be used to facilitate recognition. To this end, this letter proposes an end-to-end recognition framework in the CV domain based on the transformer model. It uses a multiscale feature extraction strategy and a CV attention mechanism to get the local and global hybrid feature. Besides, a spatiotemporal transformer (STT) block is proposed to obtain the spatiotemporal correlation between image frames to assist recognition. Finally, a residual convolutional neural network (CNN) block is introduced to promote diversity in the captured representations. In the experimental part, the recognition results of the proposed method on the real and simulated datasets are better than those of other methods. Compared with the classic sequence recognition method CV long short-term memory (CVLSTM), the recognition accuracy and kappa coefficient of the proposed method are increased by approximately 5.6% and 5.4%, respectively.

In this study, a novel method based on the random forest is presented to solve the problem of track initiation in the air‐traffic‐control (ATC) radar system. ATC radar is the most common civilian surveillance radar. There are dense targets with different moving models in its observation area. When implementing track initiation, previous heuristic methods often fail to meet its high requirements for target detection rate and false track suppression. In the method proposed in this study, the general track initiation problem is modelled as a two‐category classification problem of measurement sequences. From the historical data, the sample set is constructed. The classification function is derived by Breiman's random forest. Simulation experiments and measurement experiments show that the proposed method outperforms conventional heuristic methods in the scene of the ATC system.

Inverse synthetic aperture radar (ISAR) is an important detection approach for the classification of space targets, but the few-shot circumstances often occur due to the limitation of the imaging conditions and the large attitude changes of the maneuvering targets. At present, data augmentation and metric learning are mainly used to solve the few-shot classification problems. However, these methods are not effective when facing the space targets with large attitude changes, because they can only extract the global features and it is difficult for them to obtain an effective representation of the internal features of the image. This paper proposed a complex-valued graph classification framework, which can avoid the loss of the phase information of ISAR images. Besides, a module that can extract the rich spatial relationships between image regions and effective representations is constructed. It uses the graph information reasoning method and the Transformer structure to extract the contextual features between image regions, and overcomes the classification problems caused by large attitude changes of space targets. Furthermore, a contrast learning method is introduced to reduce the impact on classification caused by the defocusing on the images, attitude changes of targets, and imaging parameters of radars. Finally, experimental results by simulation data under different imaging parameters and laboratory measured data demonstrate that the proposed method can get more accurate results and exhibits more robustness than other few-shot learning methods for targets with large attitude changes

Inverse synthetic aperture radar (ISAR) is an effective detection method for targets. However, for the maneuvering targets, the Doppler frequency induced by an arbitrary scatterer on the target is time-varying, which will cause defocus on ISAR images and bring difficulties for the further recognition process. It is hard for traditional methods to well refocus all positions on the target well. In recent years, generative adversarial networks (GANs) achieve great success in image translation. However, the current refocusing models ignore the information of high-order terms containing in the relationship between real and imaginary parts of the data. To this end, an end-to-end refocusing network, named complex-valued pix2pixHD (CVPHD), is proposed to learn the mapping from defocus to focus, which utilizes complex-valued (CV) ISAR images as an input. A CV instance normalization layer is applied to mine the deep relationship between the complex parts by calculating the covariance of them and accelerate the training. Subsequently, an innovative adaptively weighted loss function is put forward to improve the overall refocusing effect. Finally, the proposed CVPHD is tested with the simulated and real dataset, and both can get well-refocused results. The results of comparative experiments show that the refocusing error can be reduced if extending the pix2pixHD network to the CV domain and the performance of CVPHD surpasses other autofocus methods in refocusing effects. The code and dataset have been available online ( https://github.com/yhx-hit/CVPHD ).

Signal representation is to transform the intercepted original signal into feature parameters, which is the basis for the electronic reconnaissance system to analyse and identify a radar. The traditional method is to manually design a limited number of features with prior knowledge, which can only achieve the desired results in specific applications. Therefore, it is necessary to study signal representation methods that are more adaptable, reasonable, and effective. This study proposes a feature‐level design scheme that extracts multi‐level signal feature parameters to obtain deep essential feature representations, which is consistent with deep‐learning frameworks. Through the decomposition through wavelet packet transform (WPT), the best sub‐band tree structure of the signal can be obtained, and the depth representation of the signal can be well achieved, which can meet requirements of time–frequency analysis automatically. Here, Symlets wavelet packet decomposition (WPD) is used as an example to implement multi‐layer signal feature representation. The combination of characteristic parameters at different levels can reflect the different information contained in different radar signals. The simulation results show that the method can effectively identify the modulation mode of radar signals and enhance the information mining ability from radar signals.