Despite the impressive performance of deep learning in synthetic aperture radar (SAR) automatic target recognition (ATR), its generalization capability remains a critical concern, particularly when facing domain shifts between training and testing environments. Considering the inherent robustness and interpretability of electromagnetic scattering characteristics, we explore leveraging these properties to guide deep learning training, thereby improving generalization. To this end, we propose a dual-layer dynamic scattering filtering network that leverages external physical priors to guide the learning process. The first layer adaptively generates convolutional kernels conditioned on scattering cues, enabling localized modeling of target-specific scattering phenomena. The second layer establishes a cross-domain mapping from SAR imagery to scattering features, facilitating automatic extraction of salient scattering characteristics. Furthermore, an adaptive mechanism for determining the number of scattering centers is also incorporated. Experiments conducted under significant variations between training and testing sets demonstrate that our method achieves competitive recognition accuracy while maintaining low computational cost, with only approximately 0.16M parameters and 0.002G FLOPs.

Impulse radio ultrawideband (IR-UWB) radar has the advantages of low cost, high resolution, and independence of light and weather conditions. Its potential in human activity recognition (HAR) for IoT device sensing draws interest. One challenge in this domain is effectively representing spatial static and temporal dynamic information in echo sequences. Transformers, used extensively in NLP and CV, have powerful sequence long-range dependency modeling capabilities. However, in the field of radar HAR, the application research of transformers is still insufficient. In addition, there is currently a lack of publicly available IR-UWB radar human action data sets. To this end, we proposed various fine-grained feature image calculation methods and designed an IR-UWB Radar Human Activity data set (IURHA2023). This article presents a swin transformer encoder combining cosine similarity attention and patch overlap to obtain deep spatio-temporal features of human action feature images. Compared with other proposed transformer models or traditional CNNs and RNNs, the improved swin transformer encoder performs better. To further improve the feature learning capability of the backbone network and the robustness to echo variations, we propose a supervised contrastive learning-enhanced swin transformer (SCL-SwinT). It obtains distinctions and compact embeddings by comparing the similarities of positive and negative examples partitioned according to labels. Experimental results on the IURHA2023 data set show that SCL-SwinT achieves a recognition rate exceeding 90%, and the inference speed on IoT edge devices satisfies real-time applications. Ablation experiments demonstrate the effectiveness of the proposed components. In addition, SCL-SwinT exhibits good robustness to environmental factors like noise, multipath, and distance.

We propose an innovative vision transformer (ViT)-based architecture for synthetic aperture radar (SAR) automatic target recognition (ATR), which trains models in a self-supervised learning fashion. Compared with convolution neural networks (CNNs)-based models, transformer-based architectures further focus on locational information among features, enabling models to understand images globally. However, the integral challenge with transformer is that they commonly demand more samples for training than the CNN-based models. Furthermore, securing substantial labeled SAR images is typically a daunting task, particularly for noncooperative targets. To address these issues, the proposed model combines the ViT architecture with a contrastive learning framework. The process begins by pretraining the model using substantial unlabeled samples, followed by the execution of fine-tuning with limited labeled data. Besides, A data augmentation mechanism is designed for contrastive learning to enhance diversity and amounts of samples, simultaneously learning robust representations. Experiments conducted on MSTAR datasets demonstrate that the proposed model can perform very well on SAR image classification tasks even without sufficient labeled training samples.

In this study, we introduce MobileViTX, an enhanced MobileViT architecture for human activity recognition in impulse radio ultra-wideband (IR-UWB) radar applications. MobileViT is a lightweight Vision Transformer mainly consisting of MobileViT blocks and MobileNetv2 blocks. Modifications to the MobileNetv2 block include adding a Drop Path and an SE module and altering activation functions to hard-sigmoid and hard-swish. Additionally, the self-attention in the MobileViT block is transformed to possess linear complexity. These adjustments aim to accelerate inference while preserving high accuracy. We experiment with a dataset from 20 individuals performing 20 distinct actions, using 5-fold cross-validation to assess our model’s performance. Results show MobileViTX outperforms the original MobileViT and other models in both recognition rate and efficiency.

The carrier-free ultra-wideband (UWB) radar (impulse radar) has seen a recent surge of interest. In this letter, a novel recognition system for vehicles based on the carrier-free UWB radar is proposed, in which the sparse representation is introduced as an effective feature extraction method. Based on the original K-SVD algorithm, we provide a new dictionary learning (DL) idea. Instead of only embedding discrimination criteria in the objective function, we expand and improve the optimization procedure of the K-SVD algorithm. Moreover, to alleviate the impact of the signal diversity on the recognition performance, we propose a hierarchical code constraint (HCC) and bind it to the improved K-SVD model. In this way, signals from the same class but with different distributions will be represented by the corresponding dictionary atoms. Extensive experiments prove the improved K-SVD with an HCC-IKSVD can effectively take both reconstruction capability and discriminative power of the dictionary into consideration.

Impulse radio ultrawideband (IR-UWB) radar has high range resolution, strong anti-jamming ability, and low power consumption and has been widely used in target detection and recognition. Currently, existing studies always extract artificial features of echo signals, such as time–frequency images, Doppler features, or time-domain features, and then distinguish these features through well-designed deep networks. However, these manual features are difficult to achieve task-invariant and disentangled representations. The target echo received by UWB radar also has amplitude, time-shift, and target-aspect sensitivity problems. To address the above problems, we propose a novel supervised contrastive learning (SupCon) framework to recognize different vehicles. Under label constraints, deep invariant representations are obtained through contrastive learning of echo signals, improving classification accuracy. First, a 1-D deep residual network (ResNet) is designed as the backbone, and the self-attention (SA) layer is added to extract long-range features of echo signals. Second, well-designed data augmentation methods can improve the performance of contrastive learning. Due to the integration of multiple data transformations, the model can learn invariant features by maximizing the mutual information between different signal transformations. Finally, we modify the SupCon loss function. It alleviates the conflict problem of simultaneously shrinking and expanding the distance between the positive samples in the feature space and improves the recognition performance of the model. Ablation experiments on the measured dataset show that the designed components of the method are effective. Comparative experiments on ultrawideband radar public datasets [Air Force Research Laboratory’s (AFRL) high-resolution range profile (HRRP), moving and stationary target acquisition and recognition (MSTAR)] also demonstrate the excellent classification performance of the proposed algorithm.

As a promising technique, dictionary learning (DL) for target recognition has seen a recent surge in recent years. Although many methods have been proposed to obtain discriminative dictionaries or coefficients via incorporating various constraints into the objective function, there are still two issues. First, it is well known that kinds of discriminative criteria on the objective function often involve substantial optimized items, increasing computation cost. Second, noises in the real world inevitably degrade the classification performance, while most DL algorithms disregard that. Aiming at these two problems, a hierarchical DL model is proposed for vehicle recognition based on the carrier-free ultrawideband (UWB) radar. With the purpose of successfully determining the identity of targets, we first learn several class-specific subdictionaries. Then, considering that the actual environment is filled with noises, we divided the learned dictionary atoms into signal and disturbance atoms in accordance with sparse coefficients to establish the signal dictionary and noise dictionary, respectively. Finally, the clean data are recovered over the corresponding signal dictionary, and meanwhile, the classification task is achieved. This hierarchical DL method takes into account both the noise-robust ability and discriminative power of the learned dictionary, in which the “atom selection” mechanism dramatically speeds up calculations. What is more, rather than imposing discriminative restraints on the objective function, we improve the K-SVD-based optimization process to complete hierarchical DL. Experimental results on the measured and synthetic data corroborate the effectiveness of the proposed method even under low signal-to-noise ratio (SNR) values. Especially, to testify to the generalization ability of the proposed method, we evaluate our algorithm on a public synthetic aperture radar (SAR) dataset (MSTAR).