Abstract Deep learning has revolutionized image analysis, but its applications are limited by the need for large datasets and high computational resources. Hybrid approaches that combine domain-specific, universal feature extractor with learnable neural networks offer a promising balance of efficiency and accuracy. This paper presents a hybrid model integrating a Gabor filter bank front-end with compact neural networks for efficient feature extraction and classification. Gabor filters, inherently bandpass, extract early-stage features with spatially shifted filters covering the frequency plane to balance spatial and spectral localization. We introduce separate channels capturing low- and high-frequency components to enhance feature representation while maintaining efficiency. The approach reduces trainable parameters and training time while preserving accuracy, making it suitable for resource-constrained environments. Compared to MobileNetV2 and EfficientNetB0, our model trains approximately 4–6 × faster on average while using fewer parameters and FLOPs. We compare it to pretrained networks used as feature extractors, lightweight fine-tuned models, and classical descriptors (HOG, LBP). It achieves competitive results with faster training and reduced computation. The hybrid model uses only around 0.60 GFLOPs and 0.34 M parameters, and we apply statistical significance testing (ANOVA, paired t-tests) to validate performance gains. Inference takes 0.01–0.02 s per image, up to 15 × faster than EfficientNetB0 and 8 × faster than MobileNetV2. Grad-CAM visualizations confirm localized attention on relevant regions. This work highlights integrating traditional features with deep learning to improve efficiency for resource-limited applications. Future work will address color fusion, robustness to noise, and automated filter optimization.

The conventional delay-and-sum beamforming technique results in blurred source maps due to its poor spatial resolution and high side-lobe levels. To overcome these limitations, the deconvolution approach for the mapping of acoustic sources (DAMAS) has been proposed as a postprocessing stage for image enhancement. DAMAS solves an inverse problem in the form of a system of linear equations. However, this is computationally intensive. This paper presents an approach that imposes two additional constraints to the inverse problem, namely sparsity and nonnegativity of the solution. The resulting constrained problem is solved within the Krylov projection framework. Moreover, the mapping of the sparsity penalty into the Krylov subspace is approximated by a sequence of ℓ2-norm problems via the iteratively reweighted norm (IRN) approach. Experimental results are presented which demonstrate the merits of the proposed method compared to several state-of-the-art approaches in terms of reconstruction accuracy and computation time.

The analysis of multichannel signals (MCS) has received a great deal of attention in the past few years. Modeling MCS requires depicting not only the temporal correlations within each single-channel signal (SCS) but also the interdependencies between marginal signals. The vector autoregressive (VAR) model is well adapted to providing insights to these ubiquitous dependencies, which is why it has been widely adopted for forecasting and analyzing impulse responses. Despite that, only a few studies have employed the VAR model for classification. To further explore this area, we propose a simple yet effective approach based on modeling MCS with a VAR process. To demonstrate the performance of our approach, we test it on real EEG recordings to discriminate between control and alcoholic subjects. Experimental results show that the proposed VAR approach can be very effective in MCS classification; it achieves competitive results on the benchmark dataset compared to existing state-of-the-art techniques.

The last decade has witnessed tremendous efforts to shape the Internet of things (IoT) platforms to be well suited for healthcare applications. These platforms are comprised of a network of wireless sensors to monitor several physical and physiological quantities. For instance, long-term monitoring of brain activities using wearable electroencephalogram (EEG) sensors is widely exploited in the clinical diagnosis of epileptic seizures and sleeping disorders. However, the deployment of such platforms is challenged by the high power consumption and system complexity. Energy efficiency can be achieved by exploring efficient compression techniques such as compressive sensing (CS). CS is an emerging theory that enables a compressed acquisition using well-designed sensing matrices. Moreover, system complexity can be optimized by using hardware friendly structured sensing matrices. This paper quantifies the performance of a CS-based multichannel EEG monitoring. In addition, the paper exploits the joint sparsity of multichannel EEG using subspace pursuit (SP) algorithm as well as a designed sparsifying basis in order to improve the reconstruction quality. Furthermore, the paper proposes a modification to the SP algorithm based on an adaptive selection approach to further improve the performance in terms of reconstruction quality, execution time, and the robustness of the recovery process.

In this letter, we propose a parametric modeling technique for the electrocardiogram (ECG) signal based on signal dependent orthogonal transform. The technique involves the mapping of the ECG heartbeats into the singular values (SV) domain using the left singular vectors matrix of the impulse response matrix of the LPC filter. The resulting spectral coefficients vector would be concentrated, leading to an approximation to a sum of exponentially damped sinusoids (EDS). A two-stage procedure is then used to estimate the model parameters. The Prony's method is first employed to obtain initial estimates of the model, while the Levenberg-Marquardt (LM) method is then applied to solve the non-linear least-square optimization problem. The ECG signal is reconstructed using the EDS parameters and the linear prediction coefficients via the inverse transform. The merit of the proposed modeling technique is illustrated on the clinical data collected from the MIT-BIH database including all the arrhythmias classes that are recommended by the Association for the Advancement of Medical Instrumentation (AAMI). For all the tested ECG heartbeats, the average values of the percent root mean square difference (PRDs) between the actual and the reconstructed signals were relatively low, varying between a minimum of 3.1545% for Premature Ventricular Contractions (PVC) class and a maximum of 10.8152% for Nodal Escape (NE) class.