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.
