Analyzing non-stationary respiratory sounds requires capturing brief, high-frequency acoustic transients. However, standard high-resolution time-frequency representations exhibit significant temporal redundancy and impose memory constraints for edge-based deployment, while traditional downsampling methods often omit or alias transient biomarkers. To address this redundancy-fidelity problem, we propose MelHilbertLDS, a lightweight acoustic feature representation framework. The method integrates quasi-random temporal downsampling via Low-Discrepancy Sequences (LDS) with spatial structuring using generalized Hilbert space-filling curves. LDS minimizes discrepancy bounds to provide principled coverage of transient events, acting as a temporal regularizer. Hilbert folding translates 1D temporal proximity into 2D spatial adjacency, which expands the effective receptive field of convolutional neural networks (CNNs) and compresses spatial dimensions. We evaluated the framework across three acoustic datasets, covering controlled event detection, data-scarce environments, and noisy clinical recordings. Results show that MelHilbertLDS reduces performance variance in low-resource scenarios and improves training stability. It achieves up to a 50-fold reduction in activation memory, a 44% reduction in CPU inference latency on MobileNetV3, and a 96% reduction in FLOPs on high-capacity CNNs, while maintaining competitive diagnostic performance. MelHilbertLDS offers a promising approach for automated analysis of transient acoustic events under memory constraints.
Medical image classification plays a vital role in clinical diagnosis. Although existing hybrid architectures have mitigated the challenges caused by scale variation and background interference, they remain inadequate in effectively fusing local details with global semantics. Moreover, deep feature fusion may lead to an imbalance between low-frequency structural information and high-frequency detail cues. To address these limitations, we propose DHSC-Net, a Dynamic Hybrid Network with Spectral Consistency Regularization. Specifically, we introduce a Dynamic Scale-Aware Hybrid Fusion (DSHF) module at each fusion stage to enable sample-adaptive local-global integration. Within this module, Dynamic Scale-Aware Gating (DSG) performs cross-branch dynamic weighting, while Dual-Dimension Feature Recalibration (DDFR) further refines the fused representations through sequential recalibration. Meanwhile, a spectral consistency loss is imposed during training. By explicitly decomposing fused features into low- and high-frequency components, the proposed constraint guides low-frequency structures toward the global branch and high-frequency details toward the local branch, thereby encouraging branch-wise spectral alignment in the fused features. Extensive experiments were conducted on four medical image classification datasets. The results show that DHSC-Net achieved accuracies of 91.44%, 85.78%, 94.17%, and 89.20% on Kvasir v2, ISIC 2018, COVID-19, and private UC datasets, respectively, achieving the best overall performance among the compared advanced models. Additional ablation studies and mechanism analyses support the contributions of spectral regularization and dynamic fusion design, further demonstrating the robustness and practical relevance of DHSC-Net.
This paper focuses on the time-of-arrival measurement model, and investigates the problem of robust indoor positioning in impulsive noise environments. To enhance the resistance of the conventional positioning methods to the impulsive noise, the optimum external weighting matrix is derived, and a new robust approach, which is based on the weighted ℓp-norm (WLP) minimization, is devised. To overcome the difficulty in solving the resultant nonconvex optimization problem, the majorization-minimization technique is utilized to relax it as several successive convex counterparts, which are tackled in an iterative way. Furthermore, to improve the positioning accuracy without increasing the number of base stations, a multi-snapshot framework of indoor positioning is proposed, which fuses the multiple time-of-arrival measurements to suppress impulsive noise in a statistically efficient way. The convergence, asymptotic mean square error, and computational complexity of the WLP-based approach are analyzed theoretically. Simulation results demonstrate the robustness of the proposed indoor positioning approach under the Gaussian mixture model noise by comparing with other methods.
In this paper, we present a new method for parameter estimation of multi-component linear frequency modulated (LFM) signals, using fractional Fourier transform (FrFT) and an improved super-resolution generative adversarial network (SRGAN). Different from existing estimation methods, the proposed method innovatively transforms parameter estimation into image processing, and effectively overcomes the impact of spectral overlap in FrFT domain by performing high-resolution reconstruction of signal peaks. This method initially maps multi-component LFM signals into a two-dimensional FrFT domain with coarse-step FrFT, in which each component exhibits a distinct energy peak. SRGAN-based image super-resolution (SR) is then introduced to perform denoising and high-resolution reconstruction on the low-resolution FrFT spectrum. Finally, the signal parameters are estimated using the peak locations in the high-resolution FrFT spectrum. Considering that conventional SRGAN mainly focuses on improving image perceptual quality, directly applying SR to the FrFT spectrum makes it difficult to recover peak features. To address this issue, we incorporate attention mechanisms into the network architecture to enhance signal energy concentration at peak locations. Meanwhile, an estimation loss is defined to avoid peak shifts and spurious peaks in the SR results. Extensive simulations demonstrate that the proposed method provides superior estimation accuracy and exhibits strong noise robustness.
Current sorting algorithms are fundamentally built upon Pulse Descriptor Words (PDWs), and the accuracy of PDW data critically affects subsequent sorting and identification processes. To address this limitation, this paper innovatively proposes a PDW-free joint representation learning approach based on a time-frequency point cloud transformation of directly-sampled wideband intermediate frequency (IF) data, aiming to achieve joint processing of radar signal deinterleaving and intra-pulse/inter-pulse feature extraction. Using the Short-Time Fourier Transform (STFT), the method transforms the original time-domain signal into a three-dimensional (time-frequency-amplitude) point cloud representation. Subsequently, an MT-PointMLP multi-task network is designed, incorporating a Squeeze-and-Excitation attention mechanism to adaptively calibrate channel weights. This network simultaneously learns from the point cloud and accomplishing signal sorting, intra-pulse modulation recognition, and inter-pulse repetition interval pattern classification. The proposed approach effectively bypasses the PDW measurement stage in the traditional pipeline, achieving direct feature learning and robust mapping from the raw signal to comprehensive results. Simulation demonstrate that the proposed method achieves F1-scores of 95.0%, 98.4% and 90.6% for signal sorting, intra-pulse recognition, and PRI recognition tasks, respectively, at 5 dB, indicating robust performance and generalization capability.
The Versatile Video Coding (VVC) standard achieves substantial coding gains through advanced tools such as Multiple Transform Selection (MTS), but at the cost of significantly increased encoder complexity. In particular, the transform selection module relies on a computationally intensive rate-distortion optimization (RDO) loop, which limits real-time and resource-constrained applications. This paper proposes a lightweight machine learning-based framework that fully replaces the transform RDO decision process. By learning encoder-side features and incorporating a quantization-aware modeling strategy to account for regime-dependent residual statistics, the proposed method directly predicts the optimal transform among six candidates without performing any RD cost evaluation. Experimental results show a 44.23% reduction in total encoding time while maintaining competitive coding efficiency in All-Intra configuration. Comparative evaluation results demonstrate that the proposed approach achieves significantly higher complexity reduction than existing transform-acceleration techniques. These findings confirm the feasibility of structurally substituting RD-based transform decision with lightweight supervised inference for efficient real-time VVC encoding.
Synthetic aperture radar (SAR) ship detection is crucial for maritime surveillance, yet high-precision oriented detection faces inherent challenges: boundary discontinuity and corner point disorder in Cartesian representations, background clutter, and dense target overlap with large scale variations. This paper proposes SDR-Net, a deep polar modeling network for robust oriented ship detection. First, to avoid rotation-induced representation ambiguity, we design a compact six-dimensional polar encoding scheme that represents an oriented ship by its centroid and four geometric parameters, namely the mapped polar radius, two polar angles, and mid-angle. Second, to handle dense arrangements and scale variations, we introduce a scale-based dense regression (SDR) strategy that dynamically allocates positive samples via covariance-adaptive rotated Gaussian heatmaps. Third, we design a dense regression heatmap loss (DRL) that supervises the spatial distribution of the heatmap and preserves structural information for accurate localization. Experiments on RSSDD and RSAR datasets demonstrate that SDR-Net achieves 96.8% and 93.4% mAP@0.5, respectively, outperforming the evaluated SAR ship detection and oriented object detection baselines. The source code and documentation for this work are publicly available at https://github.com/Ye-Blip/SDR-Net.
This paper proposes a framework that employs rate-splitting multiple access (RSMA), reconfigurable intelligent surfaces (RIS), and artificial jamming (AJ) to enhance physical-layer security, in which radio-frequency (RF) energy harvesting is applied. In the considered system, a source node first harvests energy from a power beacon and then transmits signals to multiple users through a RIS, while an artificial jammer is deployed to again the eavesdropper’s reception. The impact of imperfect coordination between the jammer and legitimate users is explicitly modeled to reflect practical deployment scenarios. To evaluate the secrecy performance, closed-form analytical expressions for the outage probability, intercept probability, average secrecy capacity (ASC), and secure energy efficiency (SEE) are derived for both the jamming and no-jamming cases. The optimal joint power allocation coefficients and jamming power for maximizing the ASC and SEE under different numbers of RIS elements are formulated and obtained through an optimization problem and numerical evaluations. Numerical results validate the theoretical analysis and show that the proposed framework outperforms benchmark schemes in terms of secrecy and energy efficiency, while offering design insights on improving performance through RIS size, power allocation, interference cancellation, jamming power selection, and energy-harvesting time optimization.
The safe operation of autonomous driving and Advanced Driver Assistance Systems (ADAS) depends heavily on accurate traffic sign detection. While current object detection algorithms perform well in this domain, they still struggle with small-object recognition and with insufficient multi-scale feature representation in complex scenarios, often resulting in missed or false detections. To address these issues, this paper proposes Hierarchical Multi-Scale Swin YOLO (HMS-YOLO), a hierarchical, multi-scale framework for traffic sign detection. First, Hierarchical Multi-Scale Fusion Network (HMSF-Net) is developed to build a sparse, skip-connected P2–P3–P5 feature pyramid that preserves high-frequency spatial details crucial for small objects. Dual Cross-Stage Partial with Spatial Attention (C2PSA) modules further enhance global contextual representation, while Context-Enhanced Head (CE-Head) improves multi-scale feature aggregation and localization accuracy. Moreover, Multi-Scale Swin Transformer (MS-Swin) is designed as an internal semantic refinement module in P4. It combines Multi-Scale Window Attention (MSWA) and Enhanced Feature Fusion (EFF) to jointly capture fine-grained textures and long-range dependencies. This design effectively enhances intermediate-level semantic representations. Extensive experiments on the Tsinghua-Tencent 100K (TT100K) and German Traffic Sign Detection Benchmark (GTSDB) datasets demonstrate that HMS-YOLO achieves superior detection accuracy and robustness, confirming its effectiveness for complex traffic sign detection in autonomous driving.
This paper investigates the impact of imperfect successive interference cancellation and phase misalignment errors in an active reconfigurable repeater (ARR)-aided non-orthogonal multiple access (NOMA) network and presents a one-size-fits-all framework for evaluating and improving the considered system in the context of the Internet of Things (IoT). In particular, we first develop analytical frameworks for the outage probability (OP) and ergodic capacity (EC) metrics to measure the performance of the proposed system against practical impairments. Next, we formulate and address the non-convex problems of users’ OP fairness and the system’s ergodic sum capacity maximization by optimizing the power allocation coefficient within network constraints. These optimization problems are tackled by exploiting the benefits of the developed OP and EC expressions combined with successive convex approximation techniques, which allow us to achieve near-optimal solutions in a manner that solves the expressions analytically without simulating the solution in the conventional way.
Tensor completion is an essential technique for recovering missing entries in multi-dimensional data, with applications in image restoration, video processing, and hyperspectral imaging. While low-rank tensor models have achieved notable success, under sparse sampling conditions both reconstruction quality and optimization convergence become increasingly challenging. Moreover, in many existing gradient-based tensor completion method, the Total Variation (TV) regularization only force local smoothness in latent factor space, which limits the ability to explicitly model local spatial or temporal structures in the reconstructed data. In this work, we propose TV Prox-Adan, a proximal gradient–based tensor completion algorithm designed for recovering visual data with low observed entries. The proposed method formulates tensor completion by modeling low-rank structure via Canonical Polyadic decomposition (CPD) and enforcing total variation regularization directly in the reconstructed data domain, enabling explicit preservation of local smoothness and edge structures. To minimize the resulting non-smooth and non-convex objective function, we formulate an accelerated Proximal Adan algorithm by embedding the ADAptive Nesterov momentum (Adan) optimizer into a proximal gradient scheme. Experiments on color images, videos, and multispectral images show that TV Prox-Adan achieves higher or comparable peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), along with lower relative squared error (RSE) compared with classical baselines and recent advanced methods under very low sampling ratios. At higher sampling ratios, although some competing methods may achieve slightly higher accuracy, TV Prox-Adan remains competitive across different modalities. These results indicate that TV Prox-Adan provides a robust and efficient solution for tensor completion, particularly in challenging sparse-data regimes.
RGB-Thermal semantic segmentation fuses visible texture details and infrared robustness to improve scene understanding in adverse conditions. However, existing RGB-T methods are inadequate in handling RGB feature degradation and illumination-aware fusion in extreme light. To address this issue, we propose an illumination-modality dual-awareness driven adaptive fine-grained fusion network for RGB-T semantic segmentation (IMAFNet). In this method, the illumination-aware dynamic adaptation module (IDAM) is first proposed to generate dynamic adaptive weight masks for improving the ability of light adaptation. After that, the gated attention-based modality-aware module (GAMAM) is proposed to enhance the ability of multi-modal collaborative expression. In which, GAMAM uses the dynamic adaptive weight masks generated by IDAM for modal recalibration. Then, the integrated features from each modality are split into two streams, feeding into the next encoder stage and the Feature Fusion Module (FFM), respectively. Finally, the fused features are decoded to restore spatial resolution and produce the final segmentation. Extensive experiments on two benchmark datasets of MFNet and PST900 demonstrate that the proposed IMAFNet achieves state-of-the-art performance.
In recent years, intelligent radio signal recognition has attracted increasing attention for its ability to classify signals in non-cooperative environments such as electronic reconnaissance and large-scale Internet of Things networks. In open electromagnetic environments, both known and unknown signals coexist. To enhance situation awareness and intelligence acquisition, radio signal recognition requires not only classifying known signals but also discovering unknown signals. However, existing methods are generally developed under the closed-set assumption, while directly transferring novel category discovery methods to radio signals is challenged by noisy signal representations, modulation similarity, and ambiguous feature distributions. To address this challenge, we propose a novel signal discovery method, termed prototype calibration and multi-permutation statistics (PCMS), which leverages known signals to guide the clustering of unknown signals in a unified prototype learning framework. Specifically, multi-permutation statistics hashing is used to generate reliable pairwise pseudo-labels for unlabeled unknown signals, and a prototype calibration strategy with adaptive distance loss and equiangular alignment loss progressively improves intra-class compactness and inter-class separability. Simulation results demonstrate that PCMS achieves superior classification and clustering performance in multiple scenarios compared with existing methods.
RGBT object tracking technology has made remarkable progress by integrating the complementary advantages of RGB and thermal infrared (TIR) modalities. However, how to effectively exploit the complementary information of multimodality and fully utilize the temporal context information still merits further exploration. We propose a four branch tracking framework based on cross Mamba prompt, named CMPTrack. This framework jointly accomplishes multimodal feature extraction, cross-modal interaction, and temporal context interaction, which effectively enhances the tracking robustness in complex scenarios. The cross Mamba prompt (CMP) module is the core component of CMPTrack. The cross-modal fused feature based on initial template and that based on dynamic template are as temporal context information. CMP can perform cross modeling and information fusion on the two types of context information during forward and backward scanning processes. This not only effectively captures the temporal evolution characteristics of the target appearance, but also further enhances the discriminability of the target features relative to the background features within the search region. Quantitative results on multiple benchmarks demonstrate that our method obtains superior precision and success rate. It achieves clear performance gains over traditional RGBT trackers and temporal context-based competitors. Such favorable results benefit from the joint effect of four-branch feature extraction and the temporal context interaction implemented by the CMP module.
This paper proposes an Active Channel Emphasis (ACE) module and an acoustic model termed ACE-HFAC (ACE-based Hybrid Frequency Adaptive Convolution) for polyphonic sound event detection (SED). The proposed model integrates Frequency Dynamic Convolution (FDY) and Frequency-Aware Convolution (FAC) through adaptive channel reweighting, enabling automatic selection of complementary dynamic and stationary features without relying on a fixed channel ratio. Experiments conducted on the Domestic Environment Sound Event Detection (DESED) dataset under the DCASE 2023 Task 4 setting demonstrate that ACE-HFAC achieves statistically significant improvements in temporal localization accuracy (PSDS1) over existing frequency-adaptive convolution methods, while maintaining competitive event classification performance. In particular, ACE-HFAC-384 achieves the best overall performance, with an f1-score of 52.7%, PSDS1 of 0.4438, and PSDS2 of 0.6619, while the compact ACE-HFAC-128 attains competitive results with substantially fewer parameters. In addition, Gradient-weighted Class Activation Mapping (GradCAM) based analysis validates that the ACE module adaptively emphasizes complementary feature representations conditioned on the input signal.
Ship detection plays a crucial role in the maritime domain, with Synthetic Aperture Radar (SAR) and Optical Remote Sensing (ORS) technologies providing essential support for ship monitoring. SAR provides all-weather imaging, whereas ORS excels in identifying target shapes and textures. However, existing cross-domain detection methods often focus on single-source data, limiting their effectiveness. To address this, we propose GSDL-Net (Gaussian Sparse Detail Localization-Net), a ship detection model based on cross-domain multi-source data fusion, inspired by the multimodal fusion mechanisms of the human visual cortex and Gaussian sparse localization theory. First, the Contextual Similarity-Enhanced Attention (CSE) module is designed to optimize shared features across different domains by focusing on similarities, thereby enhancing feature recognition and maintaining high precision in multi-source data fusion scenarios. Next, the Skewed Gaussian Attention (SGA) module is introduced to mitigate the dependency of self-attention weights on feature domain similarity. It improves detection precision in spatial and multi-scale domains and strengthens the learning of multi-source data fusion features. Finally, the lightweight Gaussian Detection Head for Sparse Localization (SGD) module reduces computational complexity by dynamically weighting features and using Gaussian shared convolutions, making it suitable for deployment on edge devices. Experimental results show that GSDL-Net achieves detection accuracies of 91.1% and 83.7% on SAR and ORS datasets, respectively, and 85.8% on fused multi-source datasets. The model’s parameter count (#Param) and floating-point operations (FLOPs) are reduced to 2.0M and 5.8G, respectively, highlighting its robust cross-domain performance and lightweight advantages, demonstrating the significant potential of multi-source detection models for practical applications.
Synthetic aperture radar ground moving target indication (SAR-GMTI) plays a crucial role in continuous, all-weather monitoring of complicated scenarios. However, accurately evaluating target velocities becomes highly problematic in heavily populated areas where robust and faint moving objects share a single range resolution cell. This difficulty primarily arises from energy masking phenomena and restricted Doppler resolution. Furthermore, when operating under conditions of low signal-to-noise ratio (SNR), conventional subspace techniques and Fourier transform-based approaches frequently fail to separate closely situated targets. To overcome these limitations, this study presents a novel high-resolution velocity evaluation framework for SAR-GMTI. This framework integrates an upgraded Capon spectral estimator with the Iterative Adaptive Approach (termed IAA-MCapon). By initially executing second-order phase compensation alongside range migration correction, the proposed IAA-MCapon domain successfully separates and identifies varying ground targets. Because the target spectrum preserves equal-amplitude peaks, the system's capability to discover weaker targets in crowded environments is significantly enhanced. Ultimately, the practicability and reliability of this introduced methodology are confirmed through the comprehensive analysis of both empirical measurements and simulated datasets.
Robust model fitting serves as a critical component in artificial intelligence and computer vision, extensively employed for geometric model estimation from image feature correspondences. Sampling strategy is key to robust model fitting. However, traditional algorithms like RANSAC struggle to sample all-inlier minimal subsets under high outlier ratios or multi-structure scenarios. To address this issue, a robust sampling algorithm for model fitting, termed SCUGSC, is proposed in this paper. This algorithm constructs Spatial Consensus Units (SCUs) to capture local correlations and geometric consistency, which then guide data subset selection. Candidate points with the highest consistency are selected as initial subset elements and expanded via compatibility scores. A subset refinement strategy and an SCU-based stopping criterion further enhance estimation accuracy and convergence stability under low inlier ratios. Experiments on five challenging datasets demonstrate that SCUGSC significantly outperforms seven state-of-the-art methods. SUCGSC achieves the lowest segmentation error of 0.50%, with 98.97% F-score and 98.21% precision on CIAP. Across all datasets, SCUGSC achieves an average F-score of 93.49%, substantially outperforming RANSAC, while maintaining competitive computational efficiency with runtime ranging from 0.10s to 0.25s. These results validate the effectiveness and robustness of SCUGSC in high-outlier-ratio and multi-structure scenarios.
JPEG is the most widely used image compression format on the Internet, and reversible data hiding (RDH) in JPEG images has attracted increasing attention for privacy protection. Existing two-dimensional (2D) histogram-based RDH schemes suffer from two main limitations: (1) fixed 2D mappings designed heuristically lack universality, and (2) adaptive 2D mappings often incur prohibitive computational costs. To address these issues, this paper proposes a low-complexity adaptive 2D-model-based RDH scheme for JPEG images. Zero AC coefficients are selectively used for embedding to increase capacity while controlling file size growth. A new block-sorting strategy is further designed, and the 2D model is optimized using three components: block usage length, 2D mapping, and frequency selection. By constraining the candidate space and estimating distortion through simulated embedding, the proposed scheme reduces computational complexity. Experimental results show that, on the USC-SIPI image database, the proposed scheme achieves an average PSNR of 40.39 dB and an average file size increment of 15,215 bits under QF = 60 and a payload of 12,000 bits. The proposed scheme achieves a better balance between embedding capacity, visual quality, and file size increment than state-of-the-art methods.
