29 August 2024 in IEEE Geoscience and Remote Sensing Letters
Excellent performance in supervised learning-based radar signal modulation recognition (RSMR) techniques rely on the quantity and quality of labeled datasets, while the high cost and difficulty involved in analyzing and labeling radar signal samples limits its development. A RSMR algorithm using self-supervised contrastive learning (SSCL) methodology is proposed to address this issue. Specifically, within the classical contrastive learning framework MoCo V2, a customized data augmentation method is devised to capture time-frequency features of the radar signal. Additionally, the feature extraction network ResNet50 is improved by decoupling spatial and channel filters, resulting in greater sensitivity to the time-frequency features. To enhance the recognition accuracy, two loss functions, alignment and uniformity, are employed in place of the info noise contrastive estimation (InfoNCE) loss, and both loss functions are optimized directly. The recognition accuracy of the proposed method can reach 97.66% at a signal-to-noise ratio (SNR) of 4 dB.
24 February 2025 in IEEE Transactions on Aerospace and Electronic Systems
Convolutional neural networks (CNN) have shown excellent capability in synthetic aperture radar (SAR) target recognition. However, obtaining sufficient training samples for improving network recognition is difficult due to the deficiency of labeled SAR image data. To tackle the problem, a cross-task transfer method for SAR target recognition is proposed in this paper, which is based on a designed attention multi-scale feature fusion network (AMFFNet) and an improved auxiliary classifier generative adversarial network (IACGAN). In this method, firstly, a data augmentation method based on you only cut twice (YOCT) is devised for expanding the training samples. Secondly, the partial weight parameters of AMFFNet are initialized by the pre-trained IACGAN discriminator. Finally, AMFFNet is fine-tuned using the training samples. Additionally, on the network structure of AMFFNet, multi-scale features are extracted by adopting a multi-scale depth separable convolution module, which allows the model to acquire more informative features. Attention mechanisms are employed in this network to enhance its ability to capture important features while reducing the effects of noise. A multi-level feature fusion structure is used to further raise the representational capability of the model. Experimental results on multiple open-source real SAR datasets show that the proposed method can achieve better recognition performance under different percentages of training data, and the effectiveness and superiority of the method are confirmed by its robustness to noise.
1 January 2026 in IEEE Signal Processing Letters
In synthetic aperture radar (SAR) target recognition, the interpretation of SAR images and the accurate recognition of targets are significantly affected by speckle noise. Traditional recognition methods often fail to meet the high accuracy requirements. Therefore, this letter proposes a dual-ended fusion network (DEFNet) for SAR target recognition. The network consists of three main components: the local refinement feature extraction (LRFE) network, the large-scale feature extraction (LSFE) network, and the adaptive feature interaction fusion (AFIF) module. It employs a parallel structure, utilizing the LRFE and LSFE branch modules to extract local and large-scale features, respectively. These two features are then adaptively fused through the AFIF module to further enhance feature representation. Experimental results on the moving and stationary target acquisition and recognition dataset indicate that DEFNet demonstrates significant performance improvement compared to traditional methods, showcasing its effectiveness and adaptability in SAR target recognition tasks.
4 May 2026 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
The masked autoencoder (MAE) is a self-supervised learning model that reconstructs masked patches of an input image. Its pre-trained encoder significantly enhances downstream tasks, even surpassing supervised methods. This is particularly important for synthetic aperture radar (SAR) target recognition because labeled SAR images are scarce. Leveraging large amounts of unlabeled SAR data for network pre-training mitigates this challenge. Therefore, by combining the scale-fusion spatial-channel attention (SF-SCA) and multi-head self-attention (MHSA) modules, the multi-scale perceptive convolutional transformer (MSP-CT) is proposed as the backbone of the multi-scale perceptive MAE (MSP-MAE) for SAR target recognition tasks. The MHSA module exhibits locality and sparsity in shallow features, thereby introducing redundancy in global dependency modeling. In contrast, the SF-SCA module facilitates the interaction of localized and sparse patches through a sliding window. Therefore, the MSP-CT architecture employs alternating stacks of SF-SCA and MHSA blocks to balance the ability of the model to extract local and global features. The experimental results show that MSP-MAE pre-training can effectively improve the performance of the downstream task and can greatly alleviate the problem of insufficient labeled samples.