With the progress of signal processing technology and the emergence of new system radars, the space electromagnetic environment becomes more and more complex, which puts forward higher requirements for the deinterleaving method of radar signals. Traditional signal deinterleaving algorithms rely heavily on manual experience threshold and have poor robustness. To address this problem, we designed an intelligent radar signal deinterleaving algorithm that was completed by encoding the frequency characteristic matrix and semantic segmentation network, named Sep-RefineNet. The frequency characteristic matrix can well construct the semantic features of different pulse streams of radar signals. The Sep-RefineNet semantic segmentation network can complete pixel-level segmentation of the frequency characteristic matrix and finally uses position decoding and verification to obtain the position in the original pulse stream to complete radar signals deinterleaving. The proposed method avoids the processing of threshold judgment and pulse sequence search in traditional methods. The results of the experiment show that this algorithm improves the deinterleaving accuracy and has a good against-noise ability of aliasing pulses and missing pulses.

With the development of signal processing technology and the use of new radar systems, signal aliasing and electronic interference have occurred in space. The electromagnetic signals have become extremely complicated in their current applications in space, causing difficult problems in terms of accurately identifying radar-modulated signals in low signal-to-noise ratio (SNR) environments. To address this problem, in this paper, we propose an intelligent recognition method that combines time–frequency (T–F) analysis and a deep neural network to identify radar modulation signals. The T–F analysis of the complex Morlet wavelet transform (CMWT) method is used to extract the characteristics of signals and obtain the T–F images. Adaptive filtering and morphological processing are used in T–F image enhancement to reduce the interference of noise on signal characteristics. A deep neural network with the channel-separable ResNet (Sep-ResNet) is used to classify enhanced T–F images. The proposed method completes high-accuracy intelligent recognition of radar-modulated signals in a low-SNR environment. When the SNR is −10 dB, the probability of successful recognition (PSR) is 93.44%.

Accurate work state recognition of multifunction radar (MFR) is crucial in electronic warfare, as it helps understand the enemy's intention and evaluate potential threats. A pulse‐level work state recognition method of MFR based on the residual block with spatial attention connected gated recurrent unit by features using metric coding and correlative embedding (MC‐RSG) is proposed. Metric coding is designed to generate the distance vector with time of arrival, and the correlative embedding is performed on the distance vector and raw data features to increase the feature information by extracting feature information associated with the previous and subsequent pulses in each feature sequence, respectively. Besides, we make use of the model called RSG containing the residual block with spatial attention connected gated recurrent unit to learn the features of pulse sequences and identify the radar work state label of each pulse. The experimental work shows that the method is robust and has achieved up to 97% recognition accuracy on the test dataset under ideal observation conditions and 5% higher than the comparison network in high noise observation conditions.

With the development and popularization of unmanned aerial vehicles (UAVs) and surveillance cameras, vehicle re-identification (ReID) task plays an important role in the field of urban safety. The biggest challenge in the field of vehicle ReID is how to robustly learn the common visual representation of vehicle from different viewpoints while discriminate different vehicles with similar visual appearance. In order to solve this problem, this paper designs the normalized virtual softmax loss to enlarge the inter-class distance and decrease the intra-class distance, and a vehicle ReID model is proposed by jointly training the network with the proposed loss and triplet loss. In addition, we contribute a novel UAV vehicle ReID dataset from multiple viewpoint images to verify the robustness of methods. The experimental results show that comparing with the other softmax-based losses, our method achieves better performance and gets 76.70% and 98.95% in Rank-1 on VRAI and VRAI_AIR dataset, respectively.

The metric matrix theory is an important research object of metric measure geometry and it can be used to characterize the geometric structure of a set. For intuitionistic fuzzy sets (IFS), we defined metric information matrices (MIM) of IFS by using the metric matrix theory. We introduced the Gromov–Hausdorff metric to measure the distance between any two MIMs. We then constructed a kind of metric information matrix distance knowledge measure for IFS. The proposed distance measures have the ability to measure the distance between two incomplete intuitionistic fuzzy sets. In order to reduce the information confusion caused by the disorder of MIM, we defined a homogenous metric information matrix distance by rearranging MIM. Some theorems are given to show the properties of the constructed distance measures. At the end of the paper, some numerical experiments are given to show that the proposed distances can recognize different patterns represented by IFS.

Deep neural network models are vulnerable to attacks from adversarial methods, such as gradient attacks. Evening small perturbations can cause significant differences in their predictions. Adversarial training (AT) aims to improve the model’s adversarial robustness against gradient attacks by generating adversarial samples and optimizing the adversarial training objective function of the model. Existing methods mainly focus on improving robust accuracy, balancing natural and robust accuracy and suppressing robust overfitting. They rarely consider the AT problem from the characteristics of deep neural networks themselves, such as the stability properties under certain conditions. From a mathematical perspective, deep neural networks with stable training processes may have a better ability to suppress overfitting, as their training process is smoother and avoids sudden drops in performance. We provide a proof of the existence of Ulam stability for deep neural networks. Ulam stability not only determines the existence of the solution for an operator inequality, but it also provides an error bound between the exact and approximate solutions. The feature subspace of a deep neural network with Ulam stability can be accurately characterized and constrained by a function with special properties and a controlled error boundary constant. This restricted feature subspace leads to a more stable training process. Based on these properties, we propose an adversarial training framework called Ulam stability adversarial training (US-AT). This framework can incorporate different Ulam stability conditions and benchmark AT models, optimize the construction of the optimal feature subspace, and consistently improve the model’s robustness and training stability. US-AT is simple and easy to use, and it can be easily integrated with existing multi-class AT models, such as GradAlign and TRADES. Experimental results show that US-AT methods can consistently improve the robust accuracy and training stability of benchmark models.