Wielding Neural Networks to Interpret Facial Emotions in Photographs with Fragmentary Occlusions

K. Anji; K. Sivarama; Bhanu Prakash; Bajjuri Usha; P. V. V. S.

doi:https://doi.org/10.54216/FPA.170111

Journal of Intelligent Systems and Internet of Things | 2025

BBOA-SNDAE: A Deep Learning Model for HD Prediction in Medical IoT Systems

The recent progress in the Internet of Things (IoT), Artificial Intelligence (AI), and cloud computing has revolutionized the traditional healthcare system, upgrading it into a smart healthcare system. Medical services can be enhanced by integrating essential technology such as IoT and AI. The integration of IoT and AI presents several prospects within the healthcare industry. In this research, a novel hybrid Deep Learning (DL) model called Binary Butterfly Optimization Algorithm with Stacked Non-symmetric Deep Auto-Encoder (BBOA-SNDAE) for HD (HD) prediction based on the Medical IoT technology. The key aim of the work is to categorize and predict HD utilizing clinical data with the BBOA-SDNAE model. Initially, the model is trained using the Cleveland and Statlog datasets. The input data is preprocessed and standardized utilizing the Min-Max normalization. After preprocessing, the selection of features is performed utilizing the BBOA to choose the best optimal features for improved classification. Based on the selected features, the classification is performed using the SNDAE technique. The research model was assessed based on accuracy, sensitivity, precision, specificity, NPV, and F-measure. The model attained 99.62% accuracy, 99.45% precision, 99.32% NPV, 99.56% sensitivity, 99.45% specificity, and 99.38% f-measure using the HD dataset, and the model attained 98.84% accuracy, 98.73% precision, 98.34% NPV, 98.62% sensitivity, 98.21% specificity, and 98.27% f-measure using the sensor data. The results of the research model were compared with the current model for validation, where the research model outperformed all the compared models.

groups

Radhika .B mail -

Noor Fathima mail -

Leelavathi .V .V mail -

Ambika .N .A mail -

Pratibha .S mail -

Asma Banu .S mail

link https://doi.org/10.54216/JISIoT.140105

Volume & Issue

Vol. Volume 14 / Iss. Issue 1

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2025

Intelligent Segmentor: Self Supervised Deep Learning based Multi Organ and Tumor Segmentation with Pseudo Lables Generation from CT Images

Multi Organ and tumor segmentation is the challenging task in medical imaging and surgical planning scenarios due to its diverse applications includes lesions and organs measurements and disease diagnosis respectively. Although collecting and examining labels for all classes pose severe challenges. Furthermore, Graphical Processing Unit (GPU) optimization emerge as another critical factor for multi organ and tumor segmentation. To address the mentioned conventional challenge, we designed a deep learning-based model named “Intelligent Segmentor” which performs automated segmentation in end-to-end fashion with novel semi supervised training approach. Initially, the obtained multi organ CT images is then subjected to pre-processing in terms of geometric standardization, noise removal, and intensity normalization respectively. The pre-processed image is then further provided to dual view training for effective Pseudolabel generation. The labelled data along with generated pseudolabels are provided to train the model for amplifying the model performance. After that, there are two inputs are provided to the designed segmentation model which includes dual encoders such as GoogleNet and VGG-16 for contextual and spatial information extraction in five stages, Tweaked Feature Pyramidal Network (TFPN) for dimensionality reduction and side features extraction, and Gated Fusion Module (GFM) for fusing the side features to form unified feature map. Finally, the unified feature map is the examined through convolution layers for multi organ and tumor output. We adopted FLARE 2023 dataset for validating the proposed work with existing works on 13 various organs and tumor segmentation tasks. From the results, the proposed research achieves better Dice Similarity Coefficient (DSC) and Normalized Surface Dice (NSD) through online validation and final testing than the existing works.

groups

P. Savitha mail -

Laxmi Raja mail -

R. Santhosh mail

link https://doi.org/10.54216/JISIoT.140106

Volume & Issue

Vol. Volume 14 / Iss. Issue 1

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2025

Deep Learning for Energy Forecasting Using Gated Recurrent Units and Long Short-Term Memory

Forecasting energy demand is essential for efficient grid management as it promotes steady operations, efficient markets, and sustainable energy practices. In this study, previously observed, evenly spaced energy consumption data are analysed using recurrent neural networks based on Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) architectures to extract important insights, features, and remarkable patterns. First, the study examines the influence of meteorological features on energy consumption. The most significant meteorological features are determined by computing the MIC and Pearson's correlation coefficient. The selected features are then combined with historical energy consumption data to feed the neural network. Second, to improve and optimise the performance of the proposed models, two technical indicators - the daily energy usage average and the simple moving average - are considered. The following are some instances of comparisons in terms of prediction accuracy: (1) The MAPE of the proposed model is 2.47, whereas that of the current model is 4.03. (2) The MAPE of the existing model is 25.83, whereas the proposed solution is 18.68. (3) The MAPE of the suggested model is 24.8, while the MAPE of the current model is 26.6. (4) The MAPE of the present model is 4.77, whereas the suggested approach's is 4.42.

groups

E. T. Sivadasan mail -

N. Mohana Sundaram mail -

R. Santhosh mail

link https://doi.org/10.54216/JISIoT.140107

Volume & Issue

Vol. Volume 14 / Iss. Issue 1

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2025

Modelling a Constructive Approach For Predicting Attacks Over IoT Network Environment

Internet of Things (IoT) devices are more attractive towards various vulnerable activities and nodes are easily compromised by attackers. The complexity of insecure IoT node installation relies on device heterogeneity and resource constraints because of the network ends and conventional endpoints. This work concentrates on modeling an efficient IoT-based preservation model () which is a lightweight approach used for detecting anomaly and performance various analyses at the endpoints. This work integrates linear Support Vector Machine for pattern analysis and adaptive fuzzy rule model for data pattern rule generation to examine malicious network functionality and network traffic. While adopting the rules, the compromised node needs to fulfill the generated rules; when it fails then it is considered as malicious activity. Then, the models impose network access restrictions on the compromised and terminate the further process. Thus, the nodes are prevented from further network attacks. The evaluation model is done with the use of an online available network dataset and the dataset samples are evaluated in the complex network scenario. The simulation is done in MATLAB 2020a simulation environment and the accuracy attained with this model is higher compared to other approaches. Similarly, other metrics like False Alarm Rate (FAR) are evaluated for predicting malicious network functionality. The significance of the model is evaluated based on the prediction and mitigation of various network attacks.  The anticipated model shows a prediction rate of 90.21% for DoS attacks, 89.13% for R2L, 91.65% for probe, and 93.56% for U2R attacks.

groups

B. Sowmya mail -

Nagendra Muthuluru mail

link https://doi.org/10.54216/JISIoT.140108

Volume & Issue

Vol. Volume 14 / Iss. Issue 1

Details open_in_new

Journal of Intelligent Systems and Internet of Things | 2025

Enhancing Task Scheduling Process in Fog Computing using GTO-SSSA: A Metaheuristic Approach

Task scheduling (TS) in fog computing (FC) involves efficiently allocating computing tasks to fog nodes, considering factors such as minimizing execution time, energy consumption, and latency to meet the quality-of-service (QoS) requirements of the Internet of Things (IoT) and edge devices. Efficient TS in FC is crucial for optimizing resource usage, minimizing latency, and ensuring that IoT and edge devices receive timely and high-quality services. The growing complexity of FC environments, along with the dynamic nature of IoT applications, necessitates innovative TS models using metaheuristic algorithms to allocate tasks and meet diverse quality-of-service requirements efficiently. This research introduces the GTO-SSSA (Gorilla Troops Optimization with Skip Salp Swarm Algorithm), a novel model for intelligent TS in FC environments. This model capitalizes on the collaborative nature of the GTO algorithm while incorporating enhanced exploration and exploitation capabilities via the SSSA algorithm's skipping mechanism. The primary objective of GTO-SSSA is to tackle the intricate challenges of TS in FC effectively. This includes the efficient allocation of tasks to fog nodes, considering multiple objectives such as minimizing makespan, execution time, and throughput. The GTO-SSSA model in FC demonstrates improved efficiency, consistently surpassing compared models across various task quantities with significantly reduced makespan values. Performance improvement rates for GTO-SSSA over other models show substantial gains in TS efficiency, ranging from 0.87% to 17.83%. The model exhibits scalability as it maintains its efficiency even with an increased number of tasks, aligning with the dynamic nature of IoT applications.

groups

V. Arulkumar mail -

R. Lathamanju mail -

T. Nithya mail -

T. Rajendran mail

link https://doi.org/10.54216/JISIoT.140109

Volume & Issue

Vol. Volume 14 / Iss. Issue 1

Details open_in_new

Fusion: Practice and Applications | 2025

Advanced Stress Detection and Analysis Framework using Integration of FFT, SVM, and CNN

With the prevalence of stress-related disorders on the rise, there is an increasing demand for advanced methodologies that can effectively detect and analyze stress levels. In response to this need, this research explores the integration of Fast Fourier Transform (FFT), Support Vector Machine (SVM), and Convolutional Neural Network (CNN) techniques for unlocking insights into stress dynamics from Electroencephalogram (EEG) signals. Stress, a multifaceted phenomenon with far-reaching implications for mental health, necessitates innovative approaches for its identification and management. The study begins by elucidating the complexity of stress and its impact on individuals' well-being, highlighting the urgency for accurate and efficient stress detection methodologies. Building upon this foundation, the technical intricacies of FFT, SVM, and CNN integration are explored, elucidating their respective roles in the stress detection framework. The FFT method is employed for spectral analysis of EEG signals, providing a foundation for identifying stress-related patterns in the frequency domain. The application of Artificial Neural Networks (ANNs) for feature extraction and classification is explored, leveraging their capacity to discern intricate relationships within EEG data structures. Complementing ANNs, Support Vector Machines (SVMs) are harnessed for stress level classification, capitalizing on their robustness and efficiency in handling high-dimensional data spaces. Furthermore, Convolutional Neural Networks (CNNs) are integrated into the framework to automatically learn hierarchical features from raw EEG signals, enhancing the accuracy and efficacy of stress detection methodologies. Through comprehensive evaluation and comparison with existing algorithms, the integrated approach demonstrates superior performance across key metrics. Stress detection algorithms, such as SVM, exhibit accuracy levels ranging from 70% to 96.5%, with our proposed approach achieving remarkable results. The integrated model achieves an accuracy of 96.5% and an Area under the Curve (AUC) of 0.98, surpassing existing methods in terms of accuracy, sensitivity, specificity, and AUC.

groups

V. H. Ashwin mail -

R. Jegan mail -

Subha Hency Jose mail -

P. Rajalakshmy mail -

P. Anantha Christu Raj mail

link https://doi.org/10.54216/FPA.170107

Volume & Issue

Vol. Volume 17 / Iss. Issue 1

Details open_in_new

Fusion: Practice and Applications | 2025

An Energy-Efficient Cluster-Based Routing Protocol for WBAN in Elk Herd Optimizer

A wireless body area network (WBAN) is a wireless sensor network (WSN) that is essential to monitor patient health. Sensor nodes (SNs) are commonly positioned either inside or outside the patient's body within this network. These nodes have the ability to send data to the sink node if any functional modifications in the patient are observed. Delivering efficient routing and energy management of network nodes is a complex effort in WBAN. The energy efficiency of SNs is a primary challenge to the effective deployment of WBAN. To handle this problem, a new metaheuristic optimization algorithm called Elk Herd Optimizer (EHO) is proposed in this research. This research aims to focus on energy-efficient routing methods in WBAN sensors that are connected to the human body to enhance health monitoring efficiency. The proposed WBAN model includes the deployment of eight biosensor nodes on the human body. The primary objective is to minimize the energy utilization of WBANs by selecting the most appropriate cluster heads (CHs) based on the EHO. The EHO-based routing protocol showed higher performance in WBANs in terms of energy consumption, End-to-End (E2E) delay, packet delivery rate (PDR), network lifetime (NLT), packet loss rate (PLR), and throughput. The research model was validated by comparing its findings with the existing routing protocols. The research model surpassed all the comparable models in terms of energy consumption, latency, NLT, PDR, PLR, and throughput. The routing protocol based on the EHO algorithm improves energy efficiency by effectively selecting CHs and routing paths. The EHO model efficiently reduces the total time delay, which is essential for monitoring health in real time. It achieves a high PDR while maintaining a low packet loss rate. Furthermore, the EHO-based routing extends the longevity of the network. Additionally, it enhances network performance, hence facilitating uninterrupted and dependable monitoring of health data.

groups

D. Abdul Kareem mail -

D. Rajesh mail

link https://doi.org/10.54216/FPA.170108

Volume & Issue

Vol. Volume 17 / Iss. Issue 1

Details open_in_new

Fusion: Practice and Applications | 2025

Skin Lesion Classification using Deep Learning Methods

The incidence of cancer cases has been rising rapidly over the last few decades. Skin cancer is one of the widely found types of cancer, is further classified into two main types, Melanoma and Non-Melanoma. Though Melanoma is less common than other types of skin cancer, it can be lethal if not treated promptly. But it is not the only type of skin lesion that needs attention. It becomes necessary to promptly identify and classify the skin lesions for the recovery of the patient. The machine learning models of Deep Learning prove to be very efficient in this regard. Hence, we developed a deep learning model which is an ensemble of InceptionV3, Xception and ResNet152 models. It can classify the skin lesions into seven main types -Melanoma, Melanocytic Nevi, Benign Keratosis-like lesions, Basal cell carcinoma, actinic keratosis, vascular lesions, Dermatofibroma. The method was applied to dermoscopic images from the HAM10000 dataset. The presence of noise and artifacts in the images makes it difficult to classify. So, as a preprocessing step, we performed hair removal on the dermoscopic images which is a series of methods that starts with blackhat filtering, subsequently creating a mask for inpainting and then applying the inpainting algorithm. Further Contrast enhancement was performed by applying the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm on the luminance channel of HSV image to improve the contrast of the image and also makes sure that it is not over-amplified. It is then followed by Skin Lesion Segmentation where a grabcut algorithm is applied on the enhanced image which segments the image. Thus, the segmented images are produced which are fed to the Model for training and testing. To cope up with the unbalanced dermoscopy image dataset available, we performed Image augmentation on the images generated in the previous step which alters the existing images to create some more images for the model training process, thus solving the problem of paucity of dataset and substantially increases the performance of the model. The final dataset generated is fed to the three deep learning models InceptionV3, Xception and Resnet152 which achieved an accuracy of 84.6%, 86.5% and 86.7% respectively. These were later given to two different ensemble models - Stacking and Random Forest. The Stacking model achieved an accuracy of 88.6% and Random Forest achieved an accuracy of 92.59%. The proposed system includes a GUI for a good user experience.

groups

Nyemeesha .V mail -

M. Kavitha mail

link https://doi.org/10.54216/FPA.170109

Volume & Issue

Vol. Volume 17 / Iss. Issue 1

Details open_in_new

Fusion: Practice and Applications | 2025

Real Time Sign Recognition using YOLOv8 Object Detection Algorithm for Malayalam Sign Language

Sign language recognition is important for enhancing message and user-friendliness for the community of deaf and hearing-impaired people. This paper proposes a Malayalam Sign Language (MSL) method using sign language that emerged from the state of Kerala. The main factor contributing to this emergence of such regional sign language is the absence of a standardized and consistent approach to the use of Indian Sign Language (ISL) in various states. This is due to the variations in signs, grammar, and syntax used in different regions. The system uses the You Only Look Once v8 (YOLOv8) algorithm-based object detection method which is based on Convolution Neural Network (CNN), a widely accepted deep learning neural network design employed mainly in computer vision. As the dataset for MSL is not publicly available, we used an MSL video from YouTube provided by the National Institute of Speech and Hearing for training a custom model. We pre-processed the video to extract the frames and annotate them with sign labels. Then, we trained the YOLOv8 algorithm on the annotated frames to detect the hand region and recognize signs in real time. The proposed approach achieved an accuracy of 97.21% calculated from the mean Average Precision value on the MSL dataset. The result achieved outperformed other existing approaches even while using less dataset count compared to others.

groups

Esther Daniel mail -

V. Kathiresan mail -

Priyadarshini .C mail -

Golden Nancy .R mail -

P. Sindhu mail

link https://doi.org/10.54216/FPA.170110

Volume & Issue

Vol. Volume 17 / Iss. Issue 1

Details open_in_new

Fusion: Practice and Applications | 2025

Wielding Neural Networks to Interpret Facial Emotions in Photographs with Fragmentary Occlusions

For many years, scientists have studied the way people express their emotions through body language and facial expressions. However, it is extremely difficult to accurately interpret the emotions of a person from just a single image. Interpreting facial emotions in photographs is a complex task. It is challenging to accurately detect facial emotions with the help of neural networks when the face is occluded with fragmentary blocks. With the advent of technology, emotion detection has become more accurate and reliable. It is now possible to use facial expression recognition in images to detect emotions such as happiness, sadness, anger, fear, surprise, and more. This research discusses the effectiveness of using neural networks to identify facial emotions in photographs with occlusions present. The datasets like Fer2013 dataset, CREMA-D and RAVDESS were used to train the model and the datasets were altered by implanting occlusions randomly in the images. The altered datasets were also used to evaluate the model. The challenges and opportunities that arise when neural networks are used in this context are explored. Additionally, insight is also provided into the best approach to accomplish the task.

groups

K. Anji Reddy mail -

K. Sivarama Krishna mail -

Bhanu Prakash Battula mail -

Bajjuri Usha Rani mail -

P. V. V. S. Srinivas mail

link https://doi.org/10.54216/FPA.170111

Volume & Issue

Vol. Volume 17 / Iss. Issue 1

Details open_in_new