Multichannel speech enhancement, which utilizes spatial information from microphone arrays to extract target speech from various noise, is essential for real-world applications such as hearing aids, automatic speech recognition, and audio–video conferencing. Although deep learning-based approaches have demonstrated superior performance over traditional beamformers, they are constrained by the discriminative capacity of spectral-spatial representations and high computational costs. In this paper, we present Embedding-and-Beamforming with Position-aware circular Convolution Network (EaB-ParCNet), a spatially-enhanced network under an all-neural beamforming framework that improves Embedding-and-Beamforming (EaB) architecture through a position-aware circular convolution (ParC) Network. At the core of the proposed approach, the embedding module integrates a lightweight ParC block into a U-Net architecture to learn discriminative spectral-spatial representations. The ParC block aims to effectively capture global spatial dependencies and enhance the fusion of spectral and spatial information, whereas the U-Net architecture models multi-scale contextual information and preserves fine-grained acoustic details. Experimental results on benchmark datasets demonstrate that the proposed EaB-ParCNet achieves competitive enhancement quality with a moderate increase in computational cost, offering a favorable trade-off between performance and efficiency for offline speech enhancement scenarios.
Blind image deblurring aims to simultaneously recover a sharp latent image and an unknown blur kernel from a single blurred observation, which is a typical ill-posed problem. It essentially relies on the accurate estimation of blur kernels, where salient edge information plays a critical guiding role. Traditional edge enhancement methods in the spatial domain are susceptible to noise and pseudo-edges, resulting in unreliable kernel estimation. To address this issue, a novel deblurring framework based on a dual-domain fidelity model is proposed. Explicit edge selection is achieved by employing an adaptive enhancement strategy based on wavelet transform to enhance informative high-frequency components, thus providing more robust intermediate latent images for kernel estimation. Besides, a wavelet-domain L_1 regularizer is introduced to promote kernel sparsity. Extensive experiments on synthetic and real-world benchmarks demonstrate that the proposed method achieves competitive or superior performance compared with representative state-of-the-art model-based blind image deblurring methods, in terms of both quantitative metrics (PSNR/SSIM) and visual quality. Moreover, it is particularly effective in handling large blur kernels.
Low light image enhancement (LLIE) has an important role in enhancing the visibility and reliability of images in intelligent surveillance, autonomous navigation, medical imaging, remote sensing in poor illumination. Current deep learning-based LLIE techniques, however, typically have the drawback of having to deal with large training sets, computer power and retraining if imaging conditions change. In this study, a new training-free LLIE scheme using a hybrid differential evolution–conjugate gradient optimization algorithm is proposed. Global optimization is performed using Differential Evolution with local optimization in the vicinity of the best solution obtained by DE using Conjugate Gradient to optimize 6 enhancement parameters (γ, α, β, σ, τ, δ). The optimization is guided by multiple reference- and no-reference quality metrics such as PSNR, SSIM, FSIM, LPIPS, NIQE, BRISQUE, entropy and LOE. The experiments conducted on the LOL, LOL-v2, LIME, and DICM datasets have shown better performance with the results: 25.152 dB PSNR, 0.925 SSIM and 0.917 FSIM in 46 iterations only. The efficiency, interpretability, robustness, and generalizability of the framework is also supported by the statistical analysis presented, which shows significant gains over ADE.
Traditional nonlinear active noise control (NLANC) algorithms generally employ nonlinear controllers or nonlinear secondary-path models to compensate for nonlinear distortion. However, their noise-reduction performance and adaptability may be limited when the acoustic environment changes. This paper proposes a filtered-x least mean square algorithm based on a time-domain convolutional recurrent network (TCRN-FxLMS). The proposed method alternately trains a finite impulse response (FIR) filter and a time-domain CRN to obtain an equivalent decomposition of the linear and nonlinear components of the secondary path. An inverse nonlinear CRN is then combined with an adaptive FIR filter to form a joint controller that approximately compensates for nonlinear secondary-path distortion while maintaining adaptability. Comparisons with FxLMS, THF-NLFxLMS, a FLANN-based nonlinear ANC algorithm, and Deep ANC show that the proposed method achieves better overall noise-reduction performance in most of the investigated cases. It also demonstrates stronger adaptability than Deep ANC under the investigated time-varying acoustic condition.
This paper investigates the stability of linear systems with a time-varying delay. By jointly considering the delay magnitude and the sign of its derivative, the allowable delay set is divided into four subregions. The original system is then equivalently transformed into a four-mode switched system. This construction combines refined delay-region partitioning method with switched-system modeling and enables local delay information to be exploited in each mode. A mode-dependent Lyapunov–Krasovskii functional (LKF) is constructed, in which different weighting matrices are assigned to a different mode. The average dwell-time technique is employed to handle possible jumps of the LKF at switching instants. Delay-dependent stability criteria are derived in terms of linear matrix inequalities. Two numerical examples demonstrate that the proposed method yields larger allowable delay upper bounds than several existing approaches.
