Automatic speaker identification and verification in diverse acoustic environments remains a major challenge due to the high variability of speech signals, particularly due to emotional variations and ambient noise. This paper presents a novel approach to improve the robustness and accuracy of speaker recognition systems in these complex conditions. Our method combines advanced preprocessing of audio signals with a hybrid neural network architecture, designed to efficiently capture the spatiotemporal characteristics of speech spectrograms. The proposed preprocessing process transforms audio signals of varying durations into uniform spectral representations, facilitating their analysis by deep learning models. The developed neural architecture integrates Temporal Convolutional Layers (TCN) and Long Short-Term Memory (LSTM) recurrent networks, enriched by an additive attention mechanism exploiting auxiliary information on gender and emotions. Experiments conducted on the RAVDESS and CREMA-D databases demonstrate the remarkable effectiveness of our approach. For speaker identification, the model achieves an accuracy higher than 96% on both bases. For speaker authentication, it displays a global Equal Error Rates (EER) of 1.75% for RAVDESS and 1.60% for CREMA-D, significantly outperforming some existing methods.

This paper addresses the nurse rostering problem, aiming to create an eight-week cyclic schedule that ensures an equitable distribution of work hours among nurses while adhering to a complex set of constraints. Two different modeling approaches are proposed. The initial model, incorporating strict hourly constraints, proved computationally infeasible using the CPLEX solver. By reformulating the constraints to focus on shift counts that inherently satisfy total working hours, the computational complexity was significantly reduced. The improved model efficiently provides optimal solutions on real data, demonstrating that different modeling techniques can lead to vastly different outcomes in terms of solution time and efficiency.

Online signature recognition is a prominent biometric system reliable for users’ verification that leverages the dynamic aspects of signatures. This paper presents a novel approach for dynamic signature verification and identification by combining temporal and spectral analysis techniques. The proposed method extracts discriminative features from both spatial coordinates and their derivatives, such as velocity, acceleration, and pen speed. It also applies data augmentation techniques such as rotation, scaling, and temporal distortion, to generate static images of the signatures and use spectrograms for further analysis. The experiments carried out on the SVC2004 Task 2 database, which includes genuine and skilled forgery signatures, have shown the effectiveness of the proposed method. Indeed, for verification mode, the use of Artificial neural network allows achieving a lower EER of up 2.11%, highlighting its potential for reliable signature verification. Furthermore, exploring signature images with Convolutional neural network for identification mode, allows achieving an accuracy of 99.56%. This study highlights the effectiveness of combining temporal and spectral features for signature identification and verification in security-sensitive applications.

The search for efficient organic photovoltaics materials is crucial for advancing solar energy technologies due to their potential for low-cost, lightweight, and flexible solar cells compared to traditional inorganic photovoltaics. In this study, we employed a machine learning (ML) approach to predict the key photovoltaic parameters, namely the open-circuit voltage (Voc), the short-circuit current density (Jsc) and the power conversion efficiency (PCE) of organic semiconductors used in the active-layer of organic solar cells. We trained our ML model on a comprehensive dataset of known donor-acceptor (D/A) pairs and their respective photovoltaic properties. Using this trained model, we generated and evaluated numerous novel (D/A) combinations and predicted their Voc, Jsc and PCE values. This high-throughput screening enabled us to identify promising (D/A) pairs that have not yet been explored in the literature. As a result, our findings demonstrate the power of machine learning in accelerating the discovery and optimization of new materials combinations for organic photovoltaics, potentially leading to more efficient and cost-effective solar cells, thus advancing the viability of sustainable energy solutions.

Convolutional models may effectively identify persons quickly through automatic face analysis. A convolutional neural network architecture known as the CNN cascade structure uses numerous deep convolution layers to extract hierarchical characteristics from the input image. Cascade modeling has several drawbacks, including high computational costs and complexity, difficult training, little feedback, accumulation of errors, sensitivity to model order, and challenging interpretation. In order to improve model performance due to these many issues, we propose adding a dynamic learning rate (DLR). It entails gradually modifying the learning rate in response to the model’s performance during training. The loss of training or validation error is a concern with cascade models. The CNN cascade structure and the DLR technique are combined in this paper to present a novel deep-learning methodology for reliable face recognition as a biometric security solution. The proposed cascading CNN model based on DLR was assessed on 3D face images from the MIT CBCL database. It allows achieving a higher accuracy of up to 99.65% with 5 epochs.

Palm vein recognition is a biometric method for individuals’ authentication and/or identification based on the unique patterns of veins in the palms of their hands. This paper presents a comprehensive study of the Convolutional Neural Network (CNN) for palm vein recognition. Several CNN architectures, such as VGGNet, AlexNet, and ZFNet, have been studied and adapted by proposing an improved version based on the optimization of their parameters. Two hybrid approaches based on the fusion of the improved versions of these CNN models are proposed. The proposed method was evaluated on near-infrared palm vein images from MS-PolyU database using data augmentation. The experimental results carried out have shown the high accuracy of the proposed method, allowing achieving an accuracy rate of up to 99.72%.

This paper presents a comprehensive study on deep learning for COVID-19 detection using CT-scan images. The proposed study investigates several Conventional Neural Networks (CNN) architectures such as AlexNet, ZFNet, VGGNet, and ResNet, and thus proposed a hybrid methodology base on merging the relevant optimized architectures considered for detecting COVID-19 from CT-scan images. The proposed methods have been assessed on real datasets, and the experimental results conducted have shown the effectiveness of the proposed methods, allowing achieving a higher accuracy up to 99%.