Integrating Artificial Intelligence Driven Computer Vision Framework for Enhanced Sign Language Recognition in Hearing and Speech-Impaired People

Inderjeet; P.; B.; M. V.; Naif; Elvir

doi:https://doi.org/10.54216/JISIoT.180221

Fusion: Practice and Applications | 2026

Deep Sequence Modeling of Dump Truck Sensor Data for Fuel Efficiency and Engine Health Prediction

The Fourth Industrial Revolution represents a shift to a more connected, digital world across all industries, including mining. The application of smart sensors will reduce site risks and fuel consumption, reduce equipment breakdowns, improve preventative maintenance, and improve equipment efficiency, including dump truck engines. Dump truck fuel efficiency is influenced by a number of real-world factors, including driver behavior, road and weather conditions, and vehicle specifications. Additionally, potential engine failures and other aspects can impact vehicle outcomes. By using dynamic on-road data to predict fuel consumption per trip, the industry can effectively minimize the expense associated with driving evaluations. Furthermore, analysis of data provides valuable insights into identifying the underlying causes of fuel consumption by analyzing input parameters. This paper proposes and evaluates novel models for predicting dump truck fuel consumption and engine failures in open-pit mining. These models combine the power of features derived from data collected locally by dump truck sensors and their analysis. The fuel consumption prediction architecture for open-top mining trucks using an improved Long Short-Term Memory (LSTM) model and a double-layer thick Deep Neural Network (DNN) forms the basis of the model design, which consists of two separate components. Multi-delay Recurrent Neural Network (RNN) models have been found to be efficient and accurate. The RNN architecture is applied to capture the cyclic components and complex rules in engine consumption data. This research relied on essential factors (route, vehicle speed, engine revolutions, and engine load). The proposed model outperforms existing models, achieving (MAE=0.0210), (RMSE=0.0294), (MSE=0.0009), and accuracy (R²=0.9842), demonstrating that it can produce highly accurate predictions.

groups

Raed Majeed mail -

Hiyam Hatem mail

link https://doi.org/10.54216/FPA.210204

Volume & Issue

Vol. Volume 21 / Iss. Issue 2

Details open_in_new

Fusion: Practice and Applications | 2026

An Explainable AI Fusion-Based Model for Enhanced Deepfake Detection Using Vision Transformer and InceptionResNetV1

Generative AI has made significant strides over the past few years, and this progress has accelerated the development of deepfake techniques, which can unfortunately be used for harmful purposes. It is essential to stay up-to-date with this advancement. In this paper, we present an explainable weighted average fusion deepfake detection system that combines Vision Transformer (ViT) and InceptionResNetV1 to improve classification accuracy. We also employed LIME and GradCAM++ to provide interpretability for the model decision. ViT utilizes self-attention modules to extract features, whereas InceptionResNetV1 employs convolutional layers to extract spatial features. Grad-CAM++ highlights the important regions influencing classification, and LIME examines the regional contributions. Together, these tools offer a deeper understanding of the model's decision-making process. Our fusion technique combines the outputs of both models by assigning specific weights that users can adjust interactively through the user interface. The use of these tools gives a better understanding of how the model classifies, which improves transparency and reliability in the models. The performance of the fusion strategy is tested with accuracy, precision, recall, and F1-score. Our proposed model achieves a classification accuracy of 99.19%, surpassing both ViT and InceptionResNetV1 when we evaluated them individually. To the best of our knowledge, this work represents the first deepfake detection model that combines Vision Transformer (ViT) and InceptionResNetV1 using a weighted averaging fusion approach with dual explainability techniques.

groups

Yousef A. Alsamaani mail -

Murad A. Rassam mail

link https://doi.org/10.54216/FPA.210205

Volume & Issue

Vol. Volume 21 / Iss. Issue 2

Details open_in_new

Fusion: Practice and Applications | 2026

A Hybrid Deep Learning Model Combining VGG19 and AD_Net through Feature-Level Fusion for Real-Time Skin Cancer Classification

Automated detection (AD) techniques are essential for early recognition of skin cancer. Hybrid models using feature fusion, which combine pre-trained CNNs with customized models, have shown superiority in real-time skin cancer pathology classification. This study combines VGG19 feature maps with a novel learning network based framework called AD_Net to enhance classification accuracy. VGG19 facilitated robust low-level feature extraction, while AD_Net brilliantly extracts specialized patterns. This strategy provided a flexible and fast architecture, suitable for real-time medical applications. This work led to the classification of three of the most lethal skin cancer types. The model was trained and validated using experiments on the publicly available ISIC2019 dataset. In order to improve the interpretability of the model's predictions, interpretable artificial intelligence (XAI) techniques particularly Grad-CAM were applied. Four baseline models EfficientNetB0, MobileNetV2, Inception-v3, and VGG16, were used to assess the proposal's efficacy. The suggested model outperformed the four baseline models with 99.18% accuracy, 99.0% precision, 99.0% recall, and 99.0% F1 score. Dermatologists and other medical professionals can use this method to detect skin cancer early.

groups

Ali Atshan Abdulredah mail -

Monji kherallah mail -

Faiza Charfi mail

link https://doi.org/10.54216/FPA.210206

Volume & Issue

Vol. Volume 21 / Iss. Issue 2

Details open_in_new

Fusion: Practice and Applications | 2026

Intrusion Detection and Attack Mitigation for Cloud Blade Servers via Optimized GRU Classification and Kerberos Cryptography

Cloud communication faces numerous disruptive cybersecurity threats. Various issues related to such disruption have been the subject of previous research, but detection attacks in the blade server (BS) in the cloud have not been studied. Therefore, this paper proposes an efficient intrusion detection system (IDS) framework for BS in the cloud. This framework uses Kerberos authentication-based exponential Mestre-Brainstrass curve cryptography, Sechsoftwave and sparsely centric gated recurrent unit (SSGRU). In this framework, cloud users are firstly registered to the network, and then incoming data are encrypted. The BS is then used to balance the incoming loads, and IDS is applied to detect attacks in the BS, with the data being pre-processed firstly and the big data being handled in the IDS. Afterwards, the features are extracted, from which optimal features are selected. Attacked and normal blades are classified by using the SSGRU classifier and then differentiated by generating a Sankey diagram. The attacked blades are then isolated, and the normal blades are used for load balancing on the cloud. Results indicate that this model achieved 99.43% accuracy, thus demonstrating superior performance to other models.

groups

Waleed Khalid Alzubaidi mail

link https://doi.org/10.54216/FPA.210207

Volume & Issue

Vol. Volume 21 / Iss. Issue 2

Details open_in_new

Fusion: Practice and Applications | 2026

GLU-Attention Hybrid Architecture for Dual-Biometric Passkey Generation via Neuro-Symbolic and Chaotic Dynamics

The generation of cryptographic keys from biometric traits presents an opportunity to replace traditional password-based systems with mechanisms grounded in individual physiology. Nonetheless, reliably deriving secure and reproducible keys from modalities such as fingerprints and irises remains a significant challenge, particularly under varying input conditions and constraints on entropy. In this work, we present a hybrid dual-path deep learning architecture that combines Gated Linear Units (GLUs) with Squeeze-and-Excitation (SE) modules to extract rich, multimodal embeddings from iris and fingerprint images. The model, trained on an augmented cross-modal dataset, achieved a test accuracy of 99.92% and consistently high F1-scores across 50 subjects. To derive the cryptographic key, we apply a multi-stage pipeline that blends principal component projections, distance-based feature encoding, chaotic sequence modeling based on Lorenz-like dynamics, and a lightweight error-correcting routine. These representations are fused via a custom mixing function, producing a 512-bit binary vector subsequently refined using a SHA-256-based HKDF. Evaluation of the generated keys indicates near-ideal entropy, high inter-user separation, and strong avalanche characteristics. The system also passed multiple NIST statistical randomness tests and achieved a near-zero false acceptance rate. These results support the feasibility of the proposed method for secure and repeatable biometric key generation.

groups

Nahla Abdulnabee Sameer mail -

Bashar M. Nema mail

link https://doi.org/10.54216/FPA.210208

Volume & Issue

Vol. Volume 21 / Iss. Issue 2

Details open_in_new

Fusion: Practice and Applications | 2026

Feature-Specific GAN Augmentation and Systematic Hyperparameter Optimization: A Framework for Autism Spectrum Disorder Classification

Early and accurate diagnosis of Autism Spectrum Disorder (ASD) using neuroimaging has become increasingly viable with the advent of deep learning (DL) technologies. Current clinical diagnostic processes for ASD are largely subjective and time-intensive, creating an urgent need for objective diagnostic tools. This study presents a comprehensive comparison of three prominent functional Magnetic Resonance Imaging (fMRI) feature extraction methods, ALFF (Amplitude of Low-Frequency Fluctuations), fALFF (fractional ALFF), and ReHo (Regional Homogeneity), alongside structural Magnetic Resonance Imaging  (sMRI) data, to evaluate their effectiveness in classifying ASD using various deep learning architectures. Preprocessed data from the ABIDE dataset were utilized, with uniform preprocessing pipelines applied, followed by feature extraction using the AAL (Automated Anatomical Labeling) atlas. Synthetic data augmentation was performed using Generative Adversarial Networks (GANs) to mitigate class imbalance. We trained and tuned multiple models, including 1-dimensional Convolutional Neural Networks (1D CNNs) with multi-head attention, Long Short-Term Memory (LSTM), and Vision Transformers (ViTs), with and without hyperparameter optimization. The findings indicate that the highest classification performance was attained using ALFF features with a hyperparameter-optimized CNN enhanced by attention mechanisms, achieving an accuracy of 0.83. Similarly, ReHo features yielded an equal accuracy of 0.83 when analyzed using a Vision Transformer (ViT) model. Across all experiments, functional neuroimaging features consistently outperformed structural features in classifying ASD. Notably, systematic hyperparameter tuning led to substantial improvements, particularly for ALFF-based models, where accuracy increased markedly from 59% to 83% using the CNN+Attention architecture. This study presents a comprehensive evaluation of feature types and model architectures across neuroimaging modalities, offering critical insights into their relative diagnostic value for ASD. The achieved accuracy of 83% using both ALFF and ReHo features marks a meaningful advancement in the field, setting realistic benchmarks for future research while adhering to stringent methodological rigor.

groups

Hayder M Hani mail -

Ahmed Musa Dinar mail

link https://doi.org/10.54216/FPA.210209

Volume & Issue

Vol. Volume 21 / Iss. Issue 2

Details open_in_new

Fusion: Practice and Applications | 2026

Enhancement Medical Image using U-Net Model in Three Dimensional Vitreoretinal Surgery

Vitreoretinal surgery is highly dependent on good visualization of fragile retinal surfaces for the purpose of accurate and safe operation. However, the image quality of current 3D heads-up display systems is often suboptimal, such as low contrast or inadequate sharpness, which is likely to decrease the accuracy of operation and prolong the operation duration. Improving intraoperative image quality continues to be a challenge for the advancement of the surgical results. In this paper, we advocate a deep learning-based solution to optimal imaging parameter guidance for the prospect of 3D HU-image guided VR surgery, seeking to improve vitreoretinal surface visibility during the surgery. A hybrid model that combines a U-Net-based image enhancement with a ViT for feature refinement has been learned using 212 manually optimized still frames (extracted from the ERM surgical video). The performance of the algorithm was quantitatively assessed through peak signal-to-noise ratio (PSNR) and the structural similarity index map (SSIM) and qualitatively evaluated in terms of the improvement in sharpness, brightness, and contrast. Moreover, the in-cabin usability of optimized images was investigated in an intraoperative survey. For in-vitro validation, 121 anonymous high-resolution ERM fundus images were analyzed with a 3D display coupled with the algorithm. The SSIM and PSNR of the model were 36.45±4.90 and 0.91±0.05, respectively, which indicated considerable improvements in image sharpness, brightness, and contrast. Visible ERM size and color contrast ratio were significantly enhanced in optimized images in the in-vitro studies. The results demonstrate that the developed algorithm can perform digital image enhancement effectively and has promise in the real-time applications during the 3D heads-up vitreoretinal surgeries.

groups

Shokhan M. Al-Barzinji mail -

Ahmed Abdullah Mahmood mail -

Omar Muthanna Khudhur mail -

Zaid Sami Mohsen mail

link https://doi.org/10.54216/FPA.210210

Volume & Issue

Vol. Volume 21 / Iss. Issue 2

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2026

Advanced Deep Learning Model for Image Captioning Using Customized Vision Transformer with Global Optimization Algorithm

In the image-captioning field, the excellence of produced captions is vital for the effectual interaction of visual content. Image Captioning is the main task, which unites computer vision (CV) and natural language processing (NLP), where it goals to produce graphic legends for images. A dual-fold procedure depends on precise image perception and alters language understanding both semantically and syntactically. It is gradually challenging to stay up with the modern study and consequences in image captioning owing to the developing amount of knowledge accessible on the topic.  This analysis examines into deep learning (DL) to tackle the tasks challenged by individuals with graphic impairments, targeting to improve their visual insight via advanced technologies. By tradition, the visually impaired have trusted physical support and adaptive helps for understanding and navigating visual content. With the beginning of DL, there is a unique chance to develop this scenery. In this paper, we offer an Advanced Deep Learning Method for Image Captioning Based Using Customized Transformer with a Global Optimization Algorithm (ADLIC-CTGOA). The foremost aim of ADLIC-CTGOA model is to focus on the initiation of the effectual textual image captioning of an input image. Initially, the ADLIC-CTGOA method employs preprocessing phase to enhances both image and text data: images undergo noise removal and contrast enhancement to improve quality, while text is processed by removing numbers, converting to lowercase, and text vectorization. Next, the customized swin transformer is employed for feature extraction to capture fine-grained visual features from images. In addition, the BERT Transformer model is deployed for image captioning process. To enhance the performance of proposed technique, the chaotic Aquila optimization (CAO) technique was applied for parameter tuning for enhancing the performance. A wide sort of simulation studies are executed to ensure the improved performance of ADLIC-CTGOA system. The comparative result exploration reported the betterment of the ADLIC-CTGOA model on recent approaches in terms of different evaluation measures.

groups

Suleman Alnatheer mail -

Mohammed Altaf Ahmed mail

link https://doi.org/10.54216/JISIoT.180219

Volume & Issue

Vol. Volume 18 / Iss. Issue 2

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2026

Improving Pedestrian Walkways for Individuals with Disabilities Using Heuristic Search Based Parameter Tuning with Deep Transfer Learning Models

Blind and visually challenged people face the range of practical issues by undertaking outside travels as pedestrians. In the last decade, various beneficial devices is investigated and established to assist people with disabilities move independently and safely. Anomaly detection in pedestrian paths for visually impaired individuals, using remote sensing (RS), is crucial for improving pedestrian traffic flow and safety. Engineers and investigators can create efficient methods and tools with the effect of computer vision (CV) and machine learning (ML) to recognize anomalies and alleviate possible security hazards in pedestrian walkways. With recent progress in deep learning (DL) and ML fields, researchers have realised that the image recognition problem is supposed to be developed as classification problems. This paper proposes a Coati Optimization Algorithm-Based Parameter Tuning for Pedestrian Walkways with Transfer Learning Model (COAPT-PWTLM) technique. The main goal of COAPT-PWTLM technique is to provide automatic detection of pedestrian walkways for disability using advanced models. Initially, the median filtering (MF) is employed in the image pre-processing stage to eliminate the noise from an input image data. Furthermore, the SquezeNet1.1 model is utilized for feature extraction. For the classification process, the multi-layer autoencoder (MLAE) model is implemented. Finally, the modified update coati optimization algorithm (MUCOA) model adjusts the hyperparameter range of MLAE method optimally and results in improved classification performance. The experimental validation of the COAPT-PWTLM is verified on a benchmark image dataset and the outcomes are evaluated under dissimilar measures. The experimental outcome underlined the progress of the COAPT-PWTLM model over the existing models.

groups

Reem Alshenaifi mail

link https://doi.org/10.54216/JISIoT.180220

Volume & Issue

Vol. Volume 18 / Iss. Issue 2

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2026

Integrating Artificial Intelligence Driven Computer Vision Framework for Enhanced Sign Language Recognition in Hearing and Speech-Impaired People

Sign language (SL) detection and classification for deaf persons is an essential application of machine learning (ML) and computer vision (CV) techniques. It covers emerging forms, which acquire SL implemented by entities and convert them into auditory or textual output. It is highly significant to understand that determining a correct and robust SL detection approach is a very challenging due to many tasks such as alterations in occlusions, and lighting states in hand actions and forms. Consequently, the CV and ML models is must for testing and training. A Hand gesture detection method discovers beneficial for hearing and speaking-impaired individuals by creating usage of convolutional neural network (CNN) and human-computer interface (HCI) for classifying the constant signals of SL. In this article, an Improved Fennec Fox Algorithm for Deep Learning-Based Sign Language Recognition in Hearing and Speaking Impaired People (IFFADL-SLRHSIP) technique is proposed. The presented IFFADL-SLRHSIP technique main intention is to provide effectual communication between deaf and dumb persons and normal persons utilizing CV and artificial intelligence techniques. In the IFFADL-SLRHSIP model, the enhanced SqueezeNet model is used to capture the intricate patterns and nuances of SL gestures. For detection of the SL classification process, the recurrent neural network (RNN) method is used. To optimize model performance, the improved fennec fox algorithm (IFFA) is applied for parameter tuning, enhancing the model's precision and efficiency. The experimental outputs of the IFFADL-SLRHSIP algorithm are legalized on the SL dataset. The simulation outcomes demonstrate the greater outcomes of the IFFADL-SLRHSIP approach in terms of diverse measures.

groups

Inderjeet Kaur mail -

P. Udayakumar mail -

B. Arundhati mail -

M. V. Rajesh mail -

Naif Almakayeel mail -

Elvir Akhmetshin mail

link https://doi.org/10.54216/JISIoT.180221

Volume & Issue

Vol. Volume 18 / Iss. Issue 2

Details open_in_new