Convolutional Neural Network (CNN) still suffers from overfitting problems caused by limited samples in SAR target classification. Few-shot learning (FSL) aims to learn a classifier to classify images when only a few training examples are available for each class. Recent works on FSL demonstrate that using local descriptor representations can lead to more informative representations than using image-level representations. However, local descriptors typically capture only local information in an image, while disregarding global contextual information. Therefore, relying solely on local descriptors may not fully represent the features of an image, thus affecting the classification performance of the SAR targets. To address this problem, we propose a novel global and local descriptor-based metric network (GLMnet) for few-shot SAR target classification. The proposed GLMnet consists of global-level metric relationships and local-level metric relationships. The global-level metric relationships are computed by the Brownian distance covariance metric module to capture the relationships between image blocks. On the other hand, the local-level metric relationships are generated by the image-to-class metric module. To integrate these two metric relationships, we design an adaptive metric fusion module to achieve a more discriminative measurement performance. Furthermore, to address the issue of prototypes deviating from the true distribution due to the presence of outliers or noisy samples in the BDC metric process, we design a weighted average prototype computation module to obtain reliable prototypes. Experimental results on three SAR datasets show that the proposed GLMnet can generate reliable prototypes and achieve comparable classification results in contrast to other FSL algorithms.
28 March 2024 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Synthetic aperture radar (SAR) target classification methods based on convolutional neural networks (CNN) are susceptible to overfitting due to limited samples. Additionally, the position and deformation variations of SAR targets can also affect the feature extraction capabilities of CNN. To address these issues, this paper proposes a novel few-shot learning (FSL) method, called position-aware graph neural network (PA-GNN). The PA-GNN introduces a PAM to enhance the feature representation of the target. Specifically, deformable convolution is incorporated into the embedding network to accommodate the various shapes of SAR targets. Then, the self-attention network is brought in to capture the spatial dependence of any two positions in the feature maps, and the cross-attention network calculates the cross-attention between support and query feature maps to handle the problem of unseen categories. These two attention networks can be exploited to highlight the target regions. Besides, to integrate the features extracted by these two attention networks, an adaptive feature fusion module is designed to obtain discriminative features of the target position. The PA-GNN conducts measurement by constructing multiple layers of GNN and utilizes residual learning to alleviate over-smoothing. Furthermore, we introduce a fused loss to enhance the separability between different categories and the similarity within the same category. Experiments on OpenSARShip and MSTAR datasets demonstrate that the proposed PA-GNN can achieve comparable classification results compared with other FSL algorithms.
6 March 2025 in IEEE Transactions on Instrumentation and Measurement
The down-sampling operations in convolutional neural networks (CNNs) lead to the loss of high-frequency texture features in images, posing challenges for synthetic aperture radar (SAR) automatic target recognition. To this end, some algorithms introduce the Fourier transform to retain and restore high-frequency information. However, these algorithms are prone to overfitting with a small number of SAR samples. Additionally, the similarity of different SAR target backgrounds further complicates distinguishing between targets and backgrounds during metric learning. To address these issues, a novel few-shot learning (FSL) method for SAR target classification, namely the deep Fourier-based task-aware metric network, is proposed. First, we propose a frequency domain feature extraction module. This module uses a fast Fourier transform (FFT) to convert the input image into the frequency domain, then performs convolution on the amplitude and phase information to extract and enhance features. Then, the inverse FFT and convolution operations are applied to generate the final frequency domain feature representation. Furthermore, we propose a parameter-free contrast metric module that identifies and reduces background-related features by calculating the similarity between the query sample and both the target and non-target classes. This module fully considers the context of the entire task support set, effectively improving the model’s recognition accuracy. Finally, to address the issue of increased sensitivity to noise and outliers due to unimportant local descriptors during the metric process, we design a weighted k-nearest neighbor method to enhance the model’s focus on important information, thereby improving classification accuracy. Experimental results on two public SAR datasets demonstrate that our proposed algorithm achieves higher classification accuracy than other FSL algorithms and exhibits discriminative feature representation and metric capability.
4 March 2025 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Synthetic aperture radar (SAR) automatic target recognition (ATR) methods based on convolutional neural networks require a large number of samples to achieve good generalize. However, in practical SAR applications, the scarcity of samples has always been a persistent issue, especially in the military domain, where certain target types may have very few samples. This scarcity makes SAR target feature extraction more difficult, posing a challenge to SAR ATR. To address this issue, we propose an SVD-based feature reconstruction metric network with active contrast loss for few-shot SAR target recognition. Firstly, a self-supervised despeckling module is proposed, which applies a logarithmic transformation to SAR images, converting multiplicative speckle noise into additive noise, and then uses a self-supervised denoising network to suppress Gaussian noise. Finally, the image is restored through an exponential inverse transformation, achieving effective denoising of SAR images. Secondly, an SVD-based feature reconstruction module is proposed, which performs singular value decomposition on the extracted features and retains the key singular values for reconstruction. By extracting the main structural information, noise and unimportant details are suppressed, thereby improving classification accuracy. Furthermore, an active contrastive loss is proposed to reduce the distance between features of the same class and increase the separability between features of different classes by selecting representative samples to establish positive-negative sample pairs. Experimental results demonstrate that our proposed algorithm can achieve comparable metric performance compared to other metric-based few-shot learning methods.
The consistent speckle noise in SAR images easily interferes with the semantic information of the target. Additionally, the limited quantity of supervisory information available in one-shot learning leads to poor performance. To address the aforementioned issues, we creatively propose an SAR target recognition model based on one-shot learning. This model incorporates a background noise removal technique to eliminate the interference caused by consistent speckle noise in the image. Then, a global and local complementary strategy is employed to utilize the data’s inherent a priori information as a supplement to the supervisory information. The experimental results show that our approach achieves a recognition performance of 70.867% under the three-way one-shot condition, which attains a minimum improvement of 7.467% compared to five state-of-the-art one-shot learning methods. The ablation studies demonstrate the efficacy of each design introduced in our model.
6 February 2026 in IEEE Transactions on Circuits and Systems for Video Technology
Limited by observation conditions, the availability of synthetic aperture radar (SAR) image samples is constrained, posing challenges for deep learning-based SAR target recognition. The SAR target recognition algorithms based on few-shot learning (FSL) have made significant progress, and recent studies indicate that local descriptors outperform image-level feature representations. However, the similarity in SAR target backgrounds makes some irrelevant local descriptors severely affect the accuracy of the metric. To address this issue, we propose a novel metric-based FSL framework for SAR target recognition, i.e., adaptive weighted mutual nearest neighbor network with support-query collaborative feature reconstruction. First, by proposing the support-query collaborative feature reconstruction module, calculating the similarity between the reconstructed support features and the support feature values allows the model to better capture intra-class common features and enhance intra-class consistency. Meanwhile, calculating the similarity between the reconstructed query features and the query feature values helps the model identify effective distinguishing features, highlight target saliency, and increase inter-class differentiation. Secondly, an adaptive weighted mutual nearest neighbor module is designed, where the weight of query descriptors is adjusted to highlight distinctive features and reduce background interference. Finally, a metric fusion module is proposed, which not only computes image-to-class metrics based on local descriptors but also integrates similarity from the support set to the query set as well as from the query set to the support set, enabling a discriminative similarity measure. Experimental results on three public SAR datasets demonstrate that our proposed algorithm performs better classification than other FSL algorithms.