Pedestrian attribute recognition (PAR) plays a crucial role in intelligent surveillance systems; however, existing methods often struggle to jointly model fine-grained local details and global contextual dependencies. Most approaches typically rely on either convolutional or transformer-based architectures and tend to emphasize multiscale representation or attribute-specific attention independently, resulting in suboptimal performance. To overcome these limitations, this paper proposes a hybrid PAR framework that integrates EfficientNetV2-M and Swin Transformer-Small backbones, each augmented with a multiscale feature extraction (MS) and a class-specific residual attention (CSRA) modules. A set of six model variants are developed to evaluate the contribution of these components, including two baseline models, their MS + CSRA-augmented versions, and two ensemble configurations. Experimental results on the PETA and PA-100K datasets demonstrate that the augmented models outperform their baselines, while the augmented ensemble achieves the best performance, reaching mean accuracies of 88.79% and 85.51%, respectively. Notably, the proposed augmented ensemble improves upon the previously reported mean accuracy on PA-100K by 0.59%, while maintaining competitive performance on PETA. Despite its improved accuracy, the model remains computationally efficient, achieving a throughput of 116.81 images per second, which exceeds real-time requirements of 25 to 30 frames per second for surveillance applications. Furthermore, visualization analyses demonstrate improved attention localization, supporting the interpretability of the model.

Facial attribute recognition (FAR) has garnered significant attention due to its wide-ranging applications in biometrics and security. Traditional FAR methods typically learn shared feature representations across all attributes; however, they often fail to capture the unique characteristics necessary for each attribute, thereby limiting performance. Moreover, these methods frequently neglect uncertainty quantification, which is crucial for enhancing model reliability. To address these issues, we propose a novel FAR model that integrates global and label-specific feature learning with uncertainty quantification. The proposed model utilizes EfficientNetV2B0 as the backbone architecture and introduces two specialized heads: one for refining global features through shared attention and another for learning label-specific attention. These heads were trained jointly, and their predicted probabilities were averaged during inference to improve performance. Experiments conducted on the CelebA and LFWA datasets demonstrated that the proposed model outperformed both baseline and state-of-the-art models, achieving average accuracies of 92.11% and 87.46%, respectively. Moreover, the inclusion of uncertainty quantification provided valuable insights into model confidence, which was accompanied by measurable performance improvements, with average accuracy gains of 0.01% on CelebA and 0.05% on LFWA. Despite the improvement in accuracy, the model maintained a computational efficiency of 3.1 GFLOPs and a parameter count of 24.57 million. Additionally, visualization results using Grad-CAM showed that the attention modules accurately focused on relevant facial regions, thereby validating the interpretability of the model. These results highlight the potential of our approach for accurate, efficient, and interpretable FAR in real-world applications.

With the emergence of the Internet of Things (IoT) and many smart gadgets that support artificial intelligence, it is easier than ever to acquire, reproduce, and disseminate a large number of digital data. However, these great technologies have made it possible for intruders to easily violate issues related to copyright protection, identity theft, and privacy leakage. To address such issues, several approaches have been developed, among which image watermarking has been proven to be an ideal solution. In this paper, a blind watermarking approach for RGB color images based on joint Discrete Cosine Transform (DCT) and Discrete Wavelet Transform (DWT) is proposed. First, a self-adaptive color selecting strategy is used to select either the blue or green channel of the host image for embedding purpose. Subsequently, the selected color is subdivided into non-overlapping square blocks of size 4 × 4, and then DCT is applied to each block. Afterward, the DC values obtained from each block are decomposed into four sub-bands using DWT, and then the LH middle frequency sub-band is further decomposed into four sub-bands using DWT. Lastly, the LH1 obtained from LH is utilized for watermark embedding. To provide security to the proposed approach, the watermark image is encrypted before embedding using a chaotic sequence originated from a logistic map method. Experimental results reveal that the proposed approach not only enhances watermark invisibility but also provide excellent watermark robustness, meeting the main requirements of image watermarking.

Hundreds of millions of people worldwide have recently been infected by the novel Coronavirus disease (COVID-19), causing significant damage to the health, economy, and welfare of the world's population. Moreover, the unprecedented number of patients with COVID-19 has placed a massive burden on healthcare centers, making timely and rapid diagnosis challenging. A crucial step in minimizing the impact of such problems is to automatically detect infected patients and place them under special care as quickly as possible. Deep learning algorithms, such as Convolutional Neural Networks (CNN), can be used to meet this need. Despite the desired results, most of the existing deep learning-based models were built on millions of parameters (weights), which are not applicable to devices with limited resources. Inspired by such fact, in this research, we developed two new lightweight CNN-based diagnostic models for the automatic and early detection of COVID-19 subjects from chest X-ray images. The first model was built for binary classification (COVID-19 and Normal), whereas the second one was built for multiclass classification (COVID-19, viral pneumonia, or normal). The proposed models were tested on a relatively large dataset of chest X-ray images, and the results showed that the accuracy rates of the 2- and 3-class-based classification models are 98.55% and 96.83%, respectively. The results also revealed that our models achieved competitive performance compared with the existing heavyweight models while significantly reducing cost and memory requirements for computing resources. With these findings, we can indicate that our models are helpful to clinicians in making insightful diagnoses of COVID-19 and are potentially easily deployable on devices with limited computational power and resources.

Heart Disease (HD) is often regarded as one of the deadliest human diseases. Therefore, early prediction of HD risks is crucial for prevention and treatment. Unfortunately, current clinical procedures for diagnosing HD are costly and often require an expert level of intervention. In response to this issue, researchers have recently developed various intelligent systems for the automated diagnosis of HD. Among the developed approaches, those based on artificial neural networks (ANNs) have gained more popularity due to their promising prediction results. However, to the authors’ knowledge, no research has attempted to exploit ANNs for feature extraction. Hence, research into bridging this gap is worthwhile for more excellent predictions. Motivated by this fact, this research proposes a new approach for HD prediction, utilizing a pre-trained Deep Neural Network (DNN) for feature extraction, Principal Component Analysis (PCA) for dimensionality reduction, and Logistic Regression (LR) for prediction. Cleveland, a publicly accessible HD dataset, was used to investigate the efficacy of the proposed approach (DNN + PCA + LR). Experimental results revealed that the proposed approach performs well on both the training and testing data, with accuracy rates of 91.79% and 93.33%, respectively. Furthermore, the proposed approach exhibited better performance when compared with the state-of-the-art approaches under most of the evaluation metrics used.

This paper proposes an efficient approach for person-independent facial expression recognition based on the fusion of Histogram of Oriented Gradients (HOG) descriptor and Cuttlefish Algorithm (CFA). The proposed approach employs HOG descriptor due to its outstanding performance in pattern recognition, which results in features that are robust against small local pose and illumination variations. However, it produces some irrelevant and noisy features that slow down and degrade the classification performance. To address this problem, a wrapper-based feature selector, called CFA, is used. This is because CFA is a recent bio-inspired feature selection algorithm, which has been shown to effectively select an optimal subset of features while achieving a high accuracy rate. Here, support vector machine classifier is used to evaluate the quality of the features selected by the CFA. Experimental results validated the effectiveness of the proposed approach in attaining a high recognition accuracy rate on three widely adopted datasets: CK+ (97.86%), RaFD (95.15%), and JAFFE (90.95%). Moreover, the results also indicated that the proposed approach yields competitive or even superior results compared to state-of-the-art approaches.

Over the past few decades, face recognition has become the most effective biometric technique in recognizing people's identity, as it is widely used in many areas of our daily lives. However, it is a challenging technique since facial images vary in rotations, expressions, and illuminations. To minimize the impact of these challenges, exploiting information from various feature extraction methods is recommended since one of the most critical tasks in face recognition system is the extraction of facial features. Therefore, this paper presents a new approach to face recognition based on the fusion of Gabor-based feature extraction, Fast Independent Component Analysis (Fas-tICA), and Linear Discriminant Analysis (LDA). In the presented method, first, face images are transformed to grayscale and resized to have a uniform size. After that, facial features are extracted from the aligned face image using Gabor, FastICA, and LDA methods. Finally, the nearest distance classifier is utilized to recognize the identity of the individuals. Here, the performance of six distance classifiers, namely Euclidean, Cosine, Bray-Curtis, Mahalanobis, Correlation, and Manhattan, are investigated. Experimental results revealed that the presented method attains a higher rank-one recognition rate compared to the recent approaches in the literature on four benchmarked face datasets: ORL, GT, FEI, and Yale. Moreover, it showed that the proposed method not only helps in better extracting the features but also in improving the overall efficiency of the facial recognition system.

Recent developments in the speed of the Internet and information technology have made the rapid exchange of multimedia information possible. However, these developments in technology lead to violations of information security and private information. Digital steganography provides the ability to protect private information that has become essential in the current Internet age. Among all digital media, digital video has become of interest to many researchers due to its high capacity for hiding sensitive data. Numerous video steganography methods have recently been proposed to prevent secret data from being stolen. Nevertheless, these methods have multiple issues related to visual imperceptibly, robustness, and embedding capacity. To tackle these issues, this paper proposes a new approach to video steganography based on the corner point principle and LSBs algorithm. The proposed method first uses Shi-Tomasi algorithm to detect regions of corner points within the cover video frames. Then, it uses 4-LSBs algorithm to hide confidential data inside the identified corner points. Besides, before the embedding process, the proposed method encrypts confidential data using Arnold’s cat map method to boost the security level. Experimental results revealed that the proposed method is highly secure and highly invisible, in addition to its satisfactory robustness against Salt & Pepper noise, Speckle noise, and Gaussian noise attacks, which has an average Structural Similarity Index (SSIM) of more than 0.81. Moreover, the results showed that the proposed method outperforms state-of-the-art methods in terms of visual imperceptibility, which offers excellent peak signalto-noise ratio (PSNR) of average 60.7 dB, maintaining excellent embedding capacity.

ElGamal Cryptosystem (EC) is a non-deterministic scheme which produces different outputs for the same input, making the cryptosystem more secure. On the other hand, the efficiency of its cryptosystem is low as it produces a 2:1 expansion in size from plaintext to ciphertext, resulting in a delay in execution time. Therefore, this paper presents a Modified ElGamal Cryptosystem (MEC) to increase the efficiency by speeding up the execution time and reducing the expansion rate in the file size after the encryption process. A comparison between the proposed MEC and the traditional EC is carried out using the same programming environment, and the implementation is tested using text data of different sizes. The results show that the performance of the proposed MEC is better than the traditional EC in terms of execution time and expansion rate. Whereas, the security of the proposed MEC is analogous to the traditional EC, which is based on the difficulty of solving the discrete logarithm problem.