Rapid advancements in remote sensing (RS) imaging technology have heightened the demand for the precise and efficient interpretation of large-scale, high-resolution RS images. Although segmentation algorithms based on convolutional neural networks (CNNs) or Transformers have achieved significant performance improvements, the trade-off between segmentation precision and computational complexity remains a key limitation for practical applications. Therefore, this paper proposes CVMH-UNet—a hybrid semantic segmentation network that integrates the Vision Mamba (VMamba) framework with multi-scale feature fusion—to achieve high-precision and relatively efficient RS image segmentation. CVMH-UNet comprises the following two core modules: the hybrid visual state space block (HVSSBlock) and the multi-frequency multi-scale feature fusion block (MFMSBlock). The HVSSBlock integrates convolutional branches to enhance local feature extraction while employing a cross 2D scanning method (CS2D) to capture global information from multiple directions, enabling the synergistic modeling of global and local features. The MFMSBlock introduces multi-frequency information via 2D Discrete Cosine Transform (2D DCT) and extracts multi-scale local details through point-wise convolution, thereby optimizing refined feature fusion in skip connections between the encoder and decoder. Experimental results on benchmark RS datasets demonstrate that CVMH-UNet achieves state-of-the-art segmentation accuracy with optimal computational efficiency, surpassing existing advanced methods.
8 January 2024 in IEEE Transactions on Geoscience and Remote Sensing
Synthetic aperture radar (SAR) target detectors based on deep learning have difficulty finding a good balance between accuracy and speed. Current pruning methods are usually used for backbone consistent pruning and seldom directly for the whole structure of deep learning target detectors; therefore, for edge-end applications, this article proposes a new full-link general automatic pruning algorithm for SAR target detectors, referred to as SARGap. First, SARGap automatically analyzes the network structure by creating a dependency graph, divides the pair-coupled network structure into the same group, and prunes the same channel for the same group of network structures so that the algorithm can be applied to a variety of complex target detectors. Second, an automatic pruning rate search method (APRS) is designed to search for the optimal pruning rate of each group of network structures in the target detector. Finally, to find a good balance between precision and speed in the automatic search of the pruning rate, a multiobjective optimization loss function (MOOL) is constructed as the APRS objective function. A series of experiments based on SSDD and HRSID, two large-scale SAR target detection datasets, are carried out to prove the superiority of this method. Using Yolov5s as the baseline, SARGap can compress parameters by 84.29%/82.86% and flops by 80.50%/81.93% on two datasets with almost no loss of accuracy. In addition, SARGap can be applied to any deep learning target detector and match hardware computing resources to achieve optimal full-link pruning.
With the rapid development of deep learning, significant progress has been made in remote sensing image target detection. However, methods based on deep learning are confronted with several challenges: (1) the inherent limitations of activation functions and downsampling operations in convolutional networks lead to frequency deviations and loss of local detail information, affecting fine-grained object recognition; (2) class imbalance and long-tail distributions further degrade the performance of minority categories; (3) large intra-class variations and small inter-class differences make it difficult for traditional deep learning methods to effectively extract fine-grained discriminative features. To address these issues, we propose a novel remote sensing aircraft recognition method. First, to mitigate the loss of local detail information, we introduce a learnable Gabor filter-based texture feature extractor, which enhances the discriminative feature representation of aircraft categories by capturing detailed texture information. Second, to tackle the long-tail distribution problem, we design a dynamic feature hallucination module that synthesizes diverse hallucinated samples, thereby improving the feature diversity of tail categories. Finally, to handle the challenge of large intra-class variations and small inter-class differences, we propose a contrastive learning module to enhance the spatial discriminative features of the targets. Extensive experiments on the large-scale fine-grained datasets FAIR1M and MAR20 demonstrate the effectiveness of our method, achieving detection accuracies of 53.56% and 89.72%, respectively, and surpassing state-of-the-art performance. The experimental results validate that our approach effectively addresses the key challenges in remote sensing aircraft recognition.
19 September 2024 in IEEE Transactions on Aerospace and Electronic Systems
In the context of high-dynamic electronic warfare, improving the precision of radar active jamming recognition algorithms under small sample conditions has become a research hotspot. When confronted with scarcity of labeled data and low jamming-to-noise ratio (JNR), existing intelligent models struggle to extract sufficient discriminative features, leading to suboptimal recognition accuracy. Particularly in the current environment of diverse and variable jamming types, these models exhibit inadequate robustness in jamming differentiation. To address these issues, this paper integrates transfer learning (TL) and dual graph convolutional network (DGCN) to propose a few-shot learning (FSL) method for radar active deception jamming recognition. Firstly, this method utilizes the large mini-ImageNet dataset to assist in pre-training a basic FSL model, obtaining abundant transferable general knowledge while excavating specific information about jamming from limited annotated samples through TL, thereby enhancing the generalization capability and optimizing efficiency of model. Subsequently, to further improve the robustness of model in recognizing diverse and complex jamming types, a DGCN structure is integrated into the fine-tuning of the FSL model based on deep convolutional architecture. It models the non-local correlations among jamming using a feature graph and distribution graph, facilitating the effective propagation of inter-class knowledge interaction and label information through message aggregation and propagation, making jamming features more similar within classes and more distinguishable between classes. Finally, the features extracted by the deep convolutional structure and DGCN are fused to obtain more comprehensive and enriched feature representation for accurate jamming recognition. Experimental results on typical jamming datasets demonstrate the effectiveness and superiority of the proposed recognition method under the condition of small sample and low JNRs.
23 December 2024 in IEEE Geoscience and Remote Sensing Letters
In the current fully open and highly dynamic electronic countermeasures environment, acquiring sufficient jamming label priors is particularly challenging. Especially when the labeled samples are extremely limited, existing intelligent recognition models for radar active jamming struggle to effectively learn discriminative features, resulting in inaccurate and unstable recognition results. Therefore, this letter proposes a visual-text alignment network (VTANet) that introduces the text modality to reused the label prior information, thereby leveraging labeling knowledge to enhance jamming recognition accuracy under few-shot conditions. During training, VTANet utilizes a text feature encoding module to encode text constructed based on label priors. Through a contrastive learning strategy, these obtained text features are then used to guide the visual feature encoding module to learn more discriminative visual time-frequency (TF) representations. Experimental results show that reusing label priors through the text modality significantly improves the intra-class compactness and inter-class separability of visual TF features. With the jamming-to-noise ratio of 5dB and only 2 labeled samples per jamming type, VTANet achieves a recognition accuracy of over 90%, improving by more than 5% compared to existing methods, thus demonstrating its superiority.
28 March 2025 in IEEE Transactions on Aerospace and Electronic Systems
By rapidly intercepting, storing, modulating, and forwarding a short segment of radar pulse signal, compound active deception mainlobe jamming could create preceding and lagging false targets that exhibit both deception and suppression effects with pulse compression gain, which significantly impairs the radar's ability to detect and track actual targets. This article proposes a self-supervised deep-learning-based method to automatically locate, recognize, and filter active deception jamming components. Firstly, multiple compound active jamming types that could cooperatively and efficiently exhibit both preceding and lagging false targets suppression effect against true targets energy distribution are thoroughly analyzed. Then, to achieve the maximum separability between active jamming and true targets, the short-time Fourier transform (STFT) is leveraged to fully reveal and depict the modulation texture and energy distribution of different jamming types that with various modulation modes and parameters in the 2-dimension (2-D) time-frequency (TF) spectrum domain. To mitigate the dependence on prior jamming modes and parameter information, the label-free, self-supervised feature extraction network that automatically detects and filters the multiple active jamming components in the 2-D TF spectrum domain is proposed. Both the measured and simulated jamming-contaminated echo data test results validate the effectiveness and robustness of the proposed method.