Graph matching is a cornerstone of ontology alignment, enabling semantic interoperability across heterogeneous knowledge sources. Traditional approaches, which rely on terminological, structural, or contextual similarities, often fail to capture the subtle and complex semantic relationships inherent in ontologies. In this paper, we propose a novel graph matching framework leveraging multidimensional embeddings to address these limitations. By projecting ontology entities into a high-dimensional vector space, our method encodes both structural and semantic interdependencies, enabling more accurate and efficient identification of correspondences. We introduce Graph Matching via Multidimensional Embeddings (GMME), a novel framework for ontology alignment. GMME leverages a multi-stage pipeline comprising ontology graph construction, embedding generation, similarity computation, and alignment refinement. Our framework employs adaptive thresholding and advanced similarity metrics, such as cosine similarity and scoring functions, to enhance alignment precision. The experimental investigation analyses the performance of GMME on the benchmark track supplied by the Ontology Alignment Evaluation Initiative (OAEI) and evaluates the impact of embedding dimensions on alignment accuracy. Experimental findings demonstrate the effectiveness of our GMME framework, particularly in scenarios requiring the resolution of complex, many-to-many correspondences. This work contributes to the broader domain of knowledge graph integration and provides a foundation for future research on scalable, domain-agnostic alignment techniques.

Medical ontologies have become indispensable in modern medicine, enabling the structuring of clinical knowledge, the standardization of terminology, and seamless semantic interoperability. Their extensive applications in electronic health records, clinical decision support systems, and research underscore their critical role in managing complex healthcare data. However, heterogeneity among independently developed ontologies introduces challenges that hinder efficient knowledge integration and interoperability. In this paper, we introduce a novel framework, Hierarchical Embedding for Ontology Matching (HEOM), based on hierarchical embedding techniques. The method captures structural, semantic, and contextual relationships, ensuring effective alignment of complex ontologies. By preserving hierarchical dependencies, HEOM improves the robustness and scalability of semantic interoperability in health-care systems. Our experiments, conducted using benchmark datasets from the Ontology Alignment Evaluation Initiative (OAEI), demonstrate significant improvements in precision, recall, and F-measure, highlighting the framework’s potential to advance intelligent and interoperable healthcare systems.

The 3D hand skeletal data has received considerable amount of attention due to its potential uses case for Hand Gesture Recognition (HGR) systems. Additionaly, Graph Neural Network (GNN) have been widely employed for skeletal-based HGR. Yet still, traditional models frequently suffer from inefficient feature representation and generalizability. Thus, to overcome these limitations, we introduce Spatial Graph Isomorphism Neural Networks (S-GINs), which use GIN layers to improve the feature aggregation. First, we create a graph representing skeletal-based recordings. Following that, a spatial network convolution module learns the inherent topology of hand gestures from neighbor nodes and updates it using a multilayer perceptron. This method promotes classification accuracy over other spatial based graph models including graph attention network, graph sampeling and aggregating, and graph convolutional networks. We validate S-GINs employing three benchmark datasets: MMHGD, 2MLMD, and Multi-view Leap2 , both horizontal and vertical sub-datasets. Experimental findings show that our model outperforms state-of-the-art graph based methods, by elevating the accuracies to 72%, 81%, 85%, and 83% respectively .

Global warming has caused significant thermal changes worldwide, increasing electricity demand by 20% to 30% in recent years. This surge disrupts the balance between energy production and consumption, sometimes leading to power outages or blackouts in certain countries. To mitigate this issue, optimizing HVAC systems, which account for more than 40% of global energy consumption, is a key solution, relying on either traditional or AI-based approaches. Our study examines the impact of gender, particularly women’s responses, on thermal comfort prediction in industrial environments with a high female presence. Using the ASHRAE RP-884 and Scales Project datasets and applying the XGBoost model, we demonstrated that women’s responses achieve the highest prediction accuracy at 68.60%, compared to 62.01% for men and 57.97% for combined responses. Moreover, based on the analysis of responses from different genders in the dataset excerpt, the comfort range for men falls within that of women, suggesting that integrating only female responses into thermal comfort models could enhance HVAC control, productivity, and energy efficiency.

Transfer learning is considered an effective technique that enhances model performance by using knowledge acquired from a source dataset to tackle a similar task on a target dataset. This technique is particularly valuable in fields where labeled data is limited, such as thermal comfort prediction. In the Tunisian context, the lack of specific thermal comfort data for office spaces occupied by multiple people represents a major challenge to optimizing workplace environments. To address this, we introduce a transfer learning-based approach using ASHRAE RP-884 data from countries with similar climatic conditions. In fact, these data were purified based on climate conditions and selected for office-type buildings. Three transfer learning methods were evaluated using three models: a Deep GRU-BiGRU model, a Deep GRU model, and a BiGRU model. Our results present that the transfer learning approach based on the Deep GRU-BiGRU model achieves the highest accuracy, reaching 66.15% in thermal comfort prediction, outperforming the other methods.

Vehicle Re-Identification involves identifying and matching a target vehicle with images captured from different views in a multi-camera network. This topic holds significant importance in various applications including intelligent transportation systems, video surveillance and smart city. However, Vehicle Re-Identification faces significant challenges in dynamic environments due to viewpoint variations, inter-vehicle appearance similarity, intra-class variability, illumination variation, occlusion and background clutter. To address these limitations, we propose HA-VReID, An Effective Hard Attention Model with Deep Learning for Vehicle Re-Identification that combines Hard attention mechanism for background removal and vehicle shape focus with EfficientNet-powered feature extraction for robust vehicle representation. Extensive experiments on the VeRi-776 and VRAI benchmarks demonstrate that our approach outperforms state-of-the-art methods in vehicle Re-Identification tasks.

Human Action Recognition (HAR) is integral to various domains, particularly in intelligent transportation systems where real-time understanding of driver behavior is critical for road safety. Accurately recognizing driver actions is challenging due to inter-class similarity, viewpoint variation, and real-world occlusions. This paper proposes a multi-view HAR approach utilizing a Siamese neural network architecture, fine-tuned with DenseNet201, to learn discriminative feature embeddings across front and side perspectives. Evaluations on the challenging 3MDAD dataset demonstrate the robustness of our method against viewpoint and modality shifts, achieving 58.22

Hand Gesture Recognition (HGR) achieved significant progress through diverse fields due to recent advancements in machine learning and sensor technologies. While Leap Motion Controller sensors offer convenient hand tracking and multi-modal data (skeletal and depth), the heterogeneous nature of these data modalities poses several challenges for HGR systems. In order to exploit the complementary information offered by skeleton and depth data, fusion algorithms are widely used. This paper proposes a novel Deep CNN-BiGRU model incorporating both intermediate and late fusion strategies. For each modality, we use a separate model for feature extraction step. Then, we apply fusion techniques for the decision step. Our proposed model demonstrates superior performance compared with models employed separately on skeletal or depth data, highlighting its effectiveness in exploiting the combined information for robust and accurate HGR.

The human lymphatic system is commonly affected by two primary forms of lymphoma disease: Hodgkin lymphoma and non-Hodgkin lymphoma. The second type, in particular, has emerged as a leading cause of patient mortality. Therefore, achieving a correct and early diagnosis is crucial for healthcare practitioners to devise suitable therapeutic strategies. For these reasons, in this work, We suggest the development of a computer-aided diagnosis system utilizing a novel hybrid approach that incorporates Handcrafted features and BiLSTM networks for the discrimination and analysis of patients with evolving lymphoma from those with residual masses who do not need re-treatment. Our proposed approach combines the concatenation of all extracted features obtained through various handcrafted methods (including histogram analysis, textural, and shape analysis) to analyze the functional, morphological, and anatomical aspects of each lesion. The "LWBDWMRI" databases were utilised for the experiment. We compared the experimental results of the suggested approach to each model: BiLSTM, LSTM, and VGG16. This comparison was conducted across five different approach cases: concatenating functional features only, textural features only, morphological features only, combining textural and morphological features to obtain global anatomical features, and incorporating both functional and anatomical criteria. The proposed approach achieved 96%, 97%, 98%, and 26.11 seconds for Accuracy, F1-score, Recall, and execution time, respectively.

This paper presents a novel approach for degraded face recognition using Vision Transformer (ViT) architectures. The process begins by inputting occluded face images into a Transformer encoder, which extracts discriminative features and creates an embedding vector essential for the de-occlusion process. In the next step, a Transformer decoder refines this representation by incorporating non-occluded features from the ground truth image, employing self-attention mechanisms to integrate information from both the embedding vector and ground truth features. The decoder outputs a comprehensive image of the occluded face, enriched with details from the input and ground truth images. Finally, the decoder generates a non-occluded face image, accurately reconstructing the occluded features. ViT in both encoder and decoder stages ensures efficient extraction and refinement of facial information, balancing memory and time constraints for optimal performance even in severe degradation conditions. The study uses two publicly available datasets, EKFD and IST-EURECOM LFFD. It evaluates the effectiveness of face recognition algorithms under various levels of face degradation such as partial occlusions, lighting variations, head poses, and facial expression changes.

Deploying drones with deep-learning capabilities for real-time object detection is a trendsetting strategy, particularly in dynamic environments such as Intelligent Transportation Systems. Drones, or Unmanned Aerial Vehicles, offer advantages including a wide field of view, cost-effectiveness, and operational efficiency, making them ideal for perception and surveillance in dynamic settings. However, drones face several challenges in terms of on-board computational capabilities, constraining the deployment of complex algorithms for aerial imagery analysis. In this paper, we exhaustively study the challenges of real-time object detection for drones, emphasizing the adaptability and effectiveness of deep learning-based detectors. We fine-tune and evaluate the six most used scaled versions of real-time object detectors for drone-imagery, using a benchmark, tailored for drone-captured data, to assess detection accuracy, inference speed, and computational efficiency. Our study proves that YOLOv5n excels in terms of inference speed at 454 FPS, but with lower precision, while YOLOv8n surpasses in precision but requires more resources and longer inference times. Larger versions show moderate accuracy improvement but demand increased computational resources and extended inference duration.

Person re-identification is still an open challenging task in various fields due to numerous factors, including illumination changes, background clutter, pose state variations and cloth changes. Several approaches have been suggested to address this problem in the context of deep learning. Generative models, particularly Variational Autoencoders (VAEs), have emerged as promising tools to address these challenges by learning discriminative feature representations of individual images. In this paper, we present Soft-Attention based Person Re-Identification (SAPRI), a novel approach that combines VAEs with a supervised ReID method to enhance the resilience and efficacy of ReID systems. The proposed approach focuses on data reconstruction based on soft attention. Variational autoen-coders encode principally person data, while ignoring irrelevant information. By incorporating supervised ReID, the model learns to appropriately classify persons in real world environments. Our SAPRI proposed method has been evaluated on well-known benchmarks, DukeMTMC-reID and CUHK03, demonstrating superior performance compared to existing state-of-the-art techniques in terms of the mean Average Precision evaluation metric (mAP). Additionally, qualitative results show the effectiveness of the VAE in generating discriminative representations of person images.

Real-world adversarial physical patches were shown to be successful in compromising state-of-the-art models in a variety of computer vision applications. Existing defenses that are based on either input gradient or features analysis have been compromised by recent GAN-based attacks that generate naturalistic patches. In this paper, we propose Jedi, a new defense against adversarial patches that is resilient to realistic patch attacks. Jedi tackles the patch localization problem from an information theory perspective; leverages two new ideas: (1) it improves the identification of potential patch regions using entropy analysis: we show that the entropy of adversarial patches is high, even in naturalistic patches; and (2) it improves the localization of adversarial patches, using an autoencoder that is able to complete patch regions from high entropy kernels. Jedi achieves high-precision adversarial patch localization, which we show is critical to successfully repair the images. Since Jedi relies on an input entropy analysis, it is model-agnostic, and can be applied on pre-trained off-the-shelf models without changes to the training or inference of the protected models. Jedi detects on average 90% of adversarial patches across different benchmarks and recovers up to 94% of successful patch attacks (Compared to 75% and 65% for LGS and Jujutsu, respectively).

Person Re-Identification is a critical component in modern video surveillance systems for locating individuals across cameras from various viewpoints. However, one of the significant challenges in person ReID arises when facial information is unavailable. To address this issue, we propose GF2PReID, a novel framework that leverages state-of-the-art deep learning-based ReID approaches to generate prior knowledge of the face region and provide detailed information about human body images. Our approach utilizes a deep residual network model trained with transfer learning to extract discriminative features from images with low discrimination, including low illumination and occlusion, collected from diverse datasets. Experimental results, on 2 challenging datasets: Market-1501 and CUHK03, demonstrate that our GF2PReID framework improves the datasets and significantly improves the performance of the Resnet-50 model, reaching the highest performance.

Human voice is an ideal data source for identifying people in many applications. Because of the increasing need for security in different public places, voice biometrics may be a good solution, as we can easily take voice records. This paper provides a brief overview of the approaches utilized in recognizing speakers, and then presents a novel approach for recognizing speakers in degraded smart-home conditions. The suggested approach includes a pre-processing phase, a feature extraction phase, and a classification phase, where the feature extraction phase consists of formant extraction to get the spectrum energy maxima of speech audio, dynamic time warping (DTW)to find an optimal alignment between two provided temporal sequences under definite restrictions, and refinement process to improve the results of the DTW system output. The experiments are carried out on a database containing 1,248 samples in order to validate the suggested approach. The latter has good results as regards the state of the art with 94.5% accuracy.

Since the smartphone is adapted to the ambient intelligence and the smart home systems are remotely accessed through the smartphones, there is a need for a secure authentication system based on some biometrics proprieties that can be taken from a smartphone. The identification of persons through ear and voice print is one of the basic biometric matters. The earlier research in ear recognition have shown that human ear is one of the representative human biometrics with uniqueness and stability. Indeed, the human voice is a perfect source of data for person identification in many applications. In this paper, we propose a fusion between the ear and voice biometrics in degraded conditions in a smart home context at 3 levels (feature, score, and decision). The experiments are conducted on the EVDDC database and a chimeric database (TIMIT and USTB-I). The best results are obtained with the feature level fusion (95.8%) with the KNN classifier.

Lymphoma affects the human lymphatic system, it has become one of the leading causes of patient deaths. Hence, accurate diagnosis is essential to help doctors prescribe a suitable treatment. In this work, we propose a new database that contains 50 volunteer patients treated for lymphoma on Whole Body (head/neck, chest, abdomen, and pelvis regions). These collected databases contain some MRI sequences and medical information of patients which tells whether that lesion is evolutive lymphoma or residual masses. It is mainly composed of 100000 images on axial, coronal, and sagittal plans. To highlight the utility of this dataset, we used a computer-aided diagnosis (CAD) system to recognize evolutive lymphoma from residual masses. This recognition is very important from a medical point of view, as it helps identify patients who may need additional therapy. After successful steps of database preparation, features extraction and selection, we evaluated several Machine Learning models such as Random Forest, Decision Tree, Naive Bayes, Extreme Gradient Boost, Logistic Regression, K-Nearest Neighbors (K-NN), and Support Vector Machine (SVM). The metrics of Accuracy, Precision, Sensitivity, Specificity, confusion matrix, Positive Predictive Value, Negative Predictive Value, Fl-score, missed classification, and Recall was measured to evaluate our CAD system based on AI models. The best obtained results reach 95% accuracy for SVM with RBF kernel. In addition, the proposed approach was compared to three literature works, and it gives much better results that can correctly recognize more than 64 lesions out of 67 cases. Interested researchers could contact the author to acquire the database.

Driver monitoring at nighttime is a tremendous area of research because of its crucial role to save lives and decrease traffic crashes injuries. However, this task is highly complex because of the high amount of naturalistic driving issues and the low visibility. In the gist of this paper, a novel nighttime driver action recognition network named multi-convolutional stream for hybrid network are proposed, which effectively fuses multimodal data to efficiently classify driver's actions in low visibility and a cluttered driving scene. Using the unique public driver action dataset recorded at nighttime, up to our knowledge, in two separate viewpoints, our proposed methodology beats state-of-the-art methodologies in classification performance, with an advancement of up to 10% over the best-practice methods.

This paper provides an informative overview about the structure, the operating principle, and the mathematical model of the Deep Belief Network (DBN) which is one of the best known and trusted deep learning models as well as a brief outline of the well-known and, likewise, the suggested pattern recognition applications that make wide use of this specific kind of neural network based on probability model. The application particularly targeted in this paper is the recognition of actions of violence. Hence, various experimentations have been conducted by varying several parameters using DeeBNet object-oriented MATLAB toolbox and the proposed approach returned an accuracy of 65.5 %.

The visual analysis of the electroencephalogram (EEG) is an expensive and time-consuming task. It can extract only 5% of the information held in the signal. Computer-assisted diagnosis could offer a way to obtain fast and reliable results and significantly reduce inter-and intra-assessor variability. In this document, we will present a tool for automatic analysis of EEG based on artificial neural networks. The proposed method consists in using signal processing and artificial intelligence algorithms to improve the interpretation of the EEG. For this purpose, we have two databases from the Nihon Kohden and Cadwell systems whose files are encrypted. The first step was to develop an application to decrypt and read the files. Thanks to this, the files could be decrypted in a standard format and the signals could be read. After that, we applied our method of automatic interpretation of the EEG. First, we preprocessed the signals using an Notch filter (50 Hz) and a bandpass filter (1–30Hz). Then, we extracted the features in the time-frequency domain based on three elements: the wavelet transform, its means, and its standard deviations. These features represent what we have used as inputs to our neural networks for classification. Our algorithm efficiently interpreted EEG signals with a correct classification rate of 97.9%, a sensitivity of 96.9%, and a specificity of 98.9%. These results have been deployed in an application that allows not only to visualize automatically the signals and the power spectral densities but also to extract the characteristics while displaying the wavelet transform related to the EEG signals of each chain.

While machine learning applications are getting mainstream owing to a demonstrated efficiency in solving complex problems, they suffer from inherent vulnerability to adversarial attacks. Adversarial attacks consist of additive noise to an input which can fool a detector. Recently, successful real-world printable adversarial “patches” were proven efficient against state-of-the-art neural networks. In the transition from digital noise based attacks to real-world physical attacks, the myriad of factors affecting object detection will also affect adversarial patches. Among these factors, view angle is one of the most influential, yet under-explored. In this paper, we study the effect of view angle on the effectiveness of an adversarial patch. To this aim, we propose the first approach that considers a multi-view context by combining existing adversarial patches with a perspective geometric transformation in order to simulate the effect of view angle changes. Our approach has been evaluated on two datasets: the first dataset which contains most real world constraints of a multi-view context, and the second dataset which empirically isolates the effect of view angle. The experiments show that view angle significantly affects the performance of adversarial patches, where in some cases the patch loses most of its effectiveness. We believe that these results motivate taking into account the effect of view angles in future adversarial attacks, and open up new opportunities for adversarial defenses.

Autonomous vehicles present a promising opportunity in the future of transportation systems by providing road safety. As significant progress has been made in the automatic environment perception, the detection of road obstacles remains a major challenge. Thus, to achieve reliable obstacle detection, several sensors have been employed. For short ranges, the Ultra-Wide Band (UWB) radar is utilized in order to detect objects in the near field. However, the main challenge appears in distinguishing the real target’s signature from noise in the received UWB signals. In this paper, we propose a novel framework that exploits Recurrent Neural Networks (RNNs) with UWB signals for multiple road obstacle detection. Features are extracted from the time-frequency domain using the discrete wavelet transform and are forwarded to the Long short-term memory (LSTM) network. We evaluate our approach on the OLIMP dataset which includes various driving situations with complex environment and targets from several classes. The obtained results show that the LSTM-based system outperforms the other implemented related techniques in terms of obstacle detection.

The identification of persons through partial fingerprint is one of the basic biometrie matters. The context of use is situated particularly in forensics. The main issue is the insufficiency of information contained in the partial fingerprint, it depends from the size of the proportion found in the crime scene. In this paper, we propose a novel approach for the identification of persons through partial fingerprint which uses redefined characteristics of minutiae extracted from the fingerprint image. To validate our work, we exploit two complete databases FVC 2004 and POLYU HRF to form our dataset. The proposed approach reaches recognition rates up to 98.06% and 98.82% when using datasets extracted from POLYU HRF database and FVC 2004 database respectively. It outperforms four state-of-the-art methods in term of recognition rates in different percentage of incomplete fingerprint image.

Many challenges are present in the pedestrian detection field which makes it a trending topic. Detecting pedestrian is an extremely difficult task under bad weather conditions. In order to improve and facilitate the detection task, it is required to use infra-red images. For the advanced driver-assistance systems (ADAS), more specifically those of the pedestrian detection, the camera is mounted on a moving vehicle resulting egomotion in the background. Thus another challenging problem is added. It is then required to compensate the background egomotion to obtain a background static scene. In this paper, we introduce an advanced approach for the pedestrian detection under poor weather conditions using a moving camera. First, using the interest point detector Speeded Up Robust Features (SURF), ego-motions in the background are adjusted. After that, the foreground is detected by subtracting frames. Then, a segmentation step is required to divide the images into multiple moving objects. Finally, a recognition process is applied in order to classify the moving objects into both categories: pedestrian and undefined patterns. The proposed approach was evaluated on the CVC14 dataset. Experimental results illustrate the good performance of the approach.

Our work highlights event detection system in video surveillance sequences. This should mainly distinguish acts of violence. The survey discusses the current methods and techniques that are being applied for the task of automated violence recognition in the images derived from video surveillance sequences. To do this, we propose a fusion strategy after using a variety of feature extraction algorithms to obtain the points-of-interest from input images and each of the extracted feature vectors is submitted to a classifier. In a decision fusion strategy, different classifiers are used to classify a feature vector and to establish a most suitable decision to classify the input action as violent or non-violent. We study the performance of the mentioned approaches on 21 datasets of human interaction images. Experiments were implemented in Matlab computing environment. This paper aspires to be a contribution for researchers who wish to improve the study of violent activity recognition and gather inspiration on the main challenges to tackle in this emerging field.

Over the last few years, the Electrocardiogram (ECG) was introduced as a powerful biometric modality for human authentication. Indeed, ECG has some characteristics specific to each individual. In this paper we present an authentication system based on the ECG signal. We are particularly interested in the feature extraction step where we propose new approach based on the slopes and the angles of the ECG signal. The neural network is used for the classification step. The results have been validated on a database related to 100 persons. We recorded a recognition rate (RR) equals 96.44% which is an encouraging result relative to the size of the database.