International Journal of Innovations in Science & Technology
Not a member yet
813 research outputs found
Sort by
Comparative Performance of Deep Learning Approaches for Sentiment Analysis on Pakistani Dramas and Movies Reviews
Full text linkSentiment analysis plays an important role in natural language processing, helping to understand public opinions shared through text. This study focuses on the challenge of analyzing sentiments in reviews of Pakistani dramas and movies, where mixed languages, informal expressions, and noisy data make accurate classification difficult. To solve this problem, several deep learning models were used and tested, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional LSTM (Bi-LSTM), Recurrent Neural Network (RNN), and Convolutional Neural Network (CNN). A detailed dataset of 12,000 user reviews was collected from platforms like IMDb and YouTube. The data was cleaned and prepared through steps such as tokenization, removing unnecessary columns, normalizing, and using sentiment scoring and word embedding for feature extraction. These models were trained and tested, with their performance evaluated using accuracy, precision, recall, and F1-score. Among all, the CNN model performed the best, achieving 98.71% accuracy and a 98.49% F1-score. The Bi-LSTM model was close behind, with 98.59% accuracy and a 98.47% F1-score. In the future, the research will explore the use of advanced transformer-based models like BERT and GPT for multilingual sentiment analysis. It will also aim to build real-time sentiment classification systems. Moreover, creating sentiment lexicons for regional languages and using hybrid deep learning methods are suggested to further improve accuracy and generalization
Feature-Level Fusion of CNN and Vision Transformer for Tomato Leaf Disease Identification
Full text linkTomato leaf diseases pose a serious threat to crop yield and quality, necessitating timely and accurate detection for effective management. Traditional visual inspection methods are subjective, labor-intensive, and inefficient, highlighting the need for automated solutions. This study explores the use of transfer learning and fine-tuning of deep learning models, ResNet-50 and Vision Transformers (ViT), for tomato leaf disease detection. A novel hybrid model integrating ResNet-50 and ViT through feature-level fusion is proposed to enhance classification accuracy. While ResNet-50 and ViT achieved accuracies of 95.20% and 98%, respectively, the hybrid model outperformed both with 99.07% accuracy. These results demonstrate the effectiveness and scalability of the hybrid model for early disease detection, offering a promising solution to enhance crop health and agricultural productivity
Predicting the AI-Driven Freelance Marketplace. A Case Study of Fiverr Gigs of Pakistan
Full text linkArtificial Intelligence (AI) affects worldwide economic performance because freelance marketplaces have become fundamental platforms to use AI innovations. This research examines AI effects on Pakistan\u27s freelance market using Fiverr gigs while focusing on AI mobile application development, together with AI website/software development services. The selection of these categories occurred because clients increasingly needed them, while contemporary industries required more AI-powered solutions. The research gathers Fiverr gig information systematically to identify main patterns about service types and client patterns, and freelancer earnings in the AI domains. Second-degree polynomial regression analyzed with Ridge regularization methods generated predictions about income development during 24 months from April 2025 until March 2027. Data shows a robust non-linear pattern of growth, which demonstrates that AI-powered freelance services provide substantial income benefits to Pakistani skilled freelancers, thus making Pakistan a rising center for AI-based digital freelancers development. Along with analyses of system competition, skills difficulty, and market capacity saturation, the research suggests proven methods to address these factors. This research provides strategic guidance to freelancers, together with investors and policymakers, who intend to enhance Pakistan\u27s position in the global freelance AI economy. Continuous skill development alongside enhanced infrastructure and supportive policy frameworks will enable the complete economic extraction of AI-driven freelance marketplaces, according to the research findings
A Novel Integrated Expert System Modelling Approach for Sugarcane Management
Full text linkStatistics in Pakistan show that sugarcane, cultivated in tropical and subtropical areas, produced 1.9 billion tons in 2020, achieving the highest position in the world. The existing practices and processes of sugarcane management are lacking in lack of efficiency and effectiveness, which are time-consuming and wasteful, wastage of money with improper management, creating issues of conflicts among farmers, workers, and mill administration. To overcome this significant concern, there is a dire need for an intelligent management system that could integrate the various tools, techniques, and technologies to achieve the objectives of adequate information for making rational decisions by minimizing time, cost, and optimizing the utilization of resources. The Phases of the study include: firstly, acquiring the Knowledge about the problem domain, i.e., Sugarcane Management System’s key factors, tools, and techniques, as well as SWOT (Strengths Weakness Opportunity Threat) analysis to identify the gap. In the second phase, to analyze and find priorities of the key factors and criteria weights through AHP (Analytical Hierarchy Process). Thirdly, to model the whole knowledge in different forms, like Decision Table, Weight Allocation Table, Decision Tree, and Conceptual Model etc. Finally, developing a prototype Rule-Based Expert System named ESFSMS (Expert System for Sugarcane Management System) and testing the proposed model through ESFSMS. The final report shows that the aggregate weight of all the factors equals 0.9995, which is nearly equal to 1.00, i.e., the goal. It is limited to a few factors, which can be extended in further research studies and the usage of modern techniques
Bio fusion: Advancing Biometric Authentication by Fusion of Physiological Signals
Full text linkBiometric authentication is becoming more popular due to its secure and reliable way of identifying individuals, offering clear advantages over traditional methods. Since physiological signals are unique and non-invasive, they have been widely researched for use in biometric systems. This study introduces a biometric identification system that combines machine learning with physiological signal fusion, using data from electromyography (EMG), phonocardiogram (PCG), and electrocardiogram (ECG). The data were collected from 32 participants using the BIOPAC MP-36 system. To remove power line interference and extract important frequency bands, Butterworth notch, and bandpass filters were applied to the raw signals. After pre-processing, two types of cepstral features were extracted: gamma tone cepstral coefficients (GTCCs) and Mel-frequency cepstral coefficients (MFCCs), which were analysed for their spectral properties. System performance was first tested by evaluating features from each signal individually. Then, the study examined the impact of combining pairs of signals— (ECG, PCG), (PCG, EMG), and (ECG, EMG)—using GTCC and MFCC features with different machine learning classifiers. Lastly, the GTCC and MFCC features from all three signals were combined to evaluate overall system performance. The results showed that MFCC-based features performed better than GTCC-based features for biometric authentication. The highest accuracy, 98.4%, was achieved using GTCC features with both the Fine K-nearest neighbour (KNN) and linear discriminant classifiers, while MFCC features reached 100% accuracy with the linear discriminant classifier. These findings highlight how effective cepstral features and signal fusion can be in enhancing biometric authentication performance
Trust Management for the LPWAN Devices in a Smart City
Full text linkThe Internet of Things (IoT) is used in several domains like health care, transportation, military, banking, and many more. These applications can lead to the realization of a smart city application. Recently, Low Power Wide Area Networks (LPWAN) have been getting attention to implement various IoT applications. However, LPWAN devices are deployed in an environment where they can face malicious cyber-attacks leading to compromised data. To make successful network communication, security is an important factor that must be taken into consideration. Previously, many solutions involving sophisticated data encryption and machine learning techniques have been proposed for this purpose. However, they require processing power which is mostly not available in the LPWAN devices. Here, we can apply lightweight trust management techniques to find the reliability of a node. In this article, we propose a trust management framework for securing LPWAN-based Smart City applications. Multiple Smart City case studies are considered for evaluating the proposed technique and results show better intruder detection
Tablet Guard: Load Cell based Quality Assurance with Image Processing
Full text linkEvery week, the pharmaceutical business manufactures thousands of pills, each of which must be thoroughly checked before being distributed to customers. The proposed Tablet-Guard project addresses this issue through innovative integration of multiple advanced technologies including load cell technology, artificial intelligence, and a servo motor-based removal mechanism for pharmaceutical quality assurance. The system incorporates deep learning-based image processing, coupled with a load cell using an HX711 module to inspect and assess the quality of each tablet in a blister strip as it moves along the conveyor belt. It inspects defects including irregular shapes and incomplete blister strips. The utilization of YOLOv8 enables real-time defect detection with high accuracy (mAP of 0.995), enhancing efficiency and minimizing production line disruptions. By accurately detecting and addressing defects such as broken, missing, or cracked tablets within blister strips, the system significantly minimizes the likelihood of substandard products being distributed to consumers
Design Of a Photocatalytic Reaction System for Pollutant Degradation: A Computational Study
Full text linkIn this study, Computational Fluid Dynamics (CFD) was used to model and simulate the photocatalytic degradation of methyl orange (MeO) in a stirred photoreactor, particularly in the presence of a bismuth oxide catalyst. This approach not only provides an effective method for treating wastewater by breaking down harmful dye pollutants but also highlights the potential of cost-effective and eco-friendly catalytic materials for environmental cleanup.
In the first phase, the catalyst was evenly distributed in an aqueous MeO solution, where photocatalysis was employed to degrade the pollutant. The structural properties of the catalyst were analyzed using scanning electron microscopy (SEM). Experiments were conducted to examine how different factors, such as pH and pollutant concentration, influenced MeO removal.
In the next step, CFD was used to numerically analyze MeO degradation through photocatalysis. The results showed that the photoreactor effectively broke down MeO. CFD modeling further explained the degradation mechanism, revealing that hydroxyl radicals (OH•) played a key role in the heterogeneous photocatalytic process. Photocatalysis significantly contributed to pollutant breakdown in both experimental and simulated phases.
The CFD models closely matched experimental data, confirming the findings related to fluid dynamics and species concentration. By offering deeper insights into mass transfer and reaction kinetics at a fraction of the cost and time, CFD proved to be more efficient than experimental methods in analyzing MeO degradation
Lower Limb Exo-Skeleton for Rehabilitation
Full text linkAbove-knee amputation remains a significant global issue, leaving many people physically disabled due to various natural and man-made causes, such as diseases, wars, and disasters. This article presents a novel, non-invasive active prosthesis based on electromyography (EMG). The proposed method offers a major advancement by achieving higher classification accuracy with minimal hardware requirements. Using EMG input signals, the active prosthesis controls three body postures: Sit, Stand, and Walk. These EMG signals are classified through two machine learning models: Support Vector Machine (SVM) and Long Short-Term Memory (LSTM) networks. Both models are evaluated based on accuracy. The results show that SVM outperforms LSTM, achieving a classification accuracy of 82%, while LSTM reaches 63%
Heart Sense: A novel IoT integrated Deep Learning Based ECG Image Analysis for Enhanced Heart Disease Prediction
Full text linkThe increasing advancements in the healthcare networks leveraging the unmatched capabilities of the Internet of Things for various fatal disease prediction and remote health monitoring that proved to be very beneficial in providing timely and accurate healthcare services to patients. Patients who are suffering from chronic diseases like blood pressure, kidney diseases, and heart diseases need treatment on time to avoid sudden deaths due to these ailments. To avoid this serious scenario, we have presented a novel approach for predicting heart diseases based on the Internet of Things. By leveraging the combined abilities of The IoT and Deep learning we have proposed an advanced approach that will able to predict heart diseases with increased accuracy and precision in comparison to the existing approaches along with providing timely notifications to both patients and the medical professionals to deal with the situation at hand most effectively.
We will be receiving real-time health data from the sensors which will be a wearable IoT device in our case. This collected data contains the continuously monitored information of the patient’s ECG using an ECG sensing system that is sent to the cloud for precise disease prediction. We will also be employing the patients ‘electronic health records which will contain ECG images to increase the accuracy of our results. The Deep Learning model called the transformer will be used in the proposed approach for the precise prediction of cardiovascular disease in real-time. Both the healthcare professionals and the patients are provided with the relevant information if an ailment is predicted for effective healthcare monitoring and treatment. The proposed model has better results than the existing approaches for the prediction of heart disease in terms of accuracy which is 99.8%