Proceedings of 6th International Conference on Recent Trends in Machine Learning, IoT, Smart Cities and Applications

Inhaltsverzeichnis

1. Frontmatter

2. Jaundice Stages Prediction Using Supervised Machine Learning

   M. Moorthi, Keerthipriya, Pooja, C. H. C. Alexander, Antony Athithan

   Abstract

   The purpose of the research work is to predict the level of jaundice to create an amazing and accurate model that classifies the severity of jaundice consistently with diverse standards such as bilirubin concentration, cardiac feature assessment affected person population and medical history, clinical personnel can use devices studying algorithms to correctly diagnose and classify sufferers according to specific degrees of jaundice making an allowance for well-timed intervention and personalized treatment plans. Those predictive models enhance patient care by allowing doctors to make better choices improve useful resource allocation and improve patient fitness.

3. Pleural Pathologies Detection Using Capacitive Sensors

   M. Moorthi, V. Abirami, S. Shrija, M. Madhu, Antony Athithan

   Abstract

   The condition known as pleural effusion, or “Water on the Lungs,” is characterized by an abnormal accumulation of fluid in the pleural region around the lungs. The pleura is a thin, delicate membrane that lines the inside of the chest cavity and surrounds the lungs. Their main purpose is to help and lubricate the respiratory mechanism. Normally, there is a tiny amount of fluid in this pleural region by nature, which helps the lung surfaces move smoothly during breathing. On the other hand, when pleural effusion occurs, an excessive amount of fluid accumulates in this area, which may hinder the lung’s ability to expand completely when breathing. There are several underlying reasons for this extra fluid, including infections, inflammatory diseases, heart failure, cancer, or hepatic disorders. The common symptoms of pleural effusion include cough, shortness of breath, and chest pain that gets worse as you breathe; these symptoms are frequently non-specific and can be hard to distinguish from those of the underlying illness.

4. Navigating the Complexities of Indoor Air Quality: Critical Analysis of Key Pollutants and Monitoring Sites

   Shubhi Jain, C. Balakrishnan, Raghaw Panpaliya, J. Mahalakshmi, M. Vinay, R. Gobi, Biju Kunnumpurath, Anitha Suseelan, Ajith Paul

   Abstract

   Indoor Air Quality (IAQ) plays a crucial role in human health, as individuals spend a significant portion of their time indoors. Many existing studies overlook the compounded effects of common environmental factors like temperature and humidity, as well as the health risks posed by long-term exposure to indoor pollutants. However, this review tries to explore how modern building materials, and indoor activities impact IAQ, with a focus on key pollutants such as ultrafine particulate matter (PM2.5 and smaller), volatile organic compounds ammonia, formaldehyde, chlorine, and radon. The review emphasizes the need for more research on IAQ, especially in developing regions, heritage settlements, and the need for standardized evaluation approaches. Moreover, there is a need of practical and affordable solutions that integrate advanced sensing technologies, real-time data insights, and easy-to-use systems to enable people to aware and take control of their IAQ. This review highlights the importance of more rigorous research, creative strategies, and strong policy support to address these challenges, particularly for communities in developing regions, crowded spaces, and vulnerable populations. The main motive is to enhance understanding of indoor air pollution and promote the development of healthier, more sustainable indoor environments. The objective is to support SDG-3 Good Health and Well Being, SDG-11 Sustainable Cities and Communities, SDG-13 Climate Action.

5. Automated Irrigation and Hazard Detection Using IoT for Sustainable Farming

   Vanshika Nigam, Tanisha Sharma, Satyajit Pangaonkar, Rushikesh Pawar, Surendra Rahamatkar

   Abstract

   The agricultural sector faces growing challenges in resource management and environmental monitoring, especially with the increasing need for sustainable and efficient farming practices. Traditional irrigation methods are often labor-intensive and prone to inefficiencies, leading to water wastage and reduced crop yields. Recent advancements in IoT have opened new possibilities for automating these processes, enabling precise control over environmental variables. Prior studies have explored various IoT applications in agriculture, such as automated irrigation systems and real-time environmental monitoring, underscoring the potential of these technologies to enhance productivity and sustainability. Building on this foundation, this investigation introduces an IoT-based elegant system for plant care designed to optimize irrigation and monitor environmental threats. The system employs the ESP32 microcontroller with sensors for moisture, ultrasonic, flame, and flow detection, automating critical agricultural tasks. Through integration with the Blynk app, the system provides real-time data to farmers, allowing remote monitoring and control. The outcomes of the system reveal significant improvements in water conservation and efficiency in crop management. Precision irrigation, guided by soil moisture readings, minimizes water wastage, while the flow sensor ensures optimal usage. Hazard detection features, including animal intrusion and fire alerts, provide timely interventions, reducing potential crop damage. Field tests show a marked reduction in manual labor and an increase in crop yield, demonstrating the system’s potential to transform small to medium-sized farms. This research not only offers a scalable, cost-effective solution but also contributes to the growing body of literature on the application of IoT in sustainable agriculture.

6. A Deep Attention Network Based Approach for Healthcare Provider Fraud Detection

   Morarjee Kolla, Jangam Niharika, S. China Ramu

   Abstract

   The integrity and financial stability of healthcare systems around the world are seriously threatened by healthcare fraud. It includes dishonest tactics including upcoding treatments to more costly alternatives, invoicing for services that were never provided, and falsifying patient diagnoses. These dishonest practices compromise the standard and reliability of patient care in addition to causing large financial losses. Conventional machine learning models and rule-based techniques are two examples of traditional fraud detection systems that frequently fail to adequately handle this issue. Low accuracy and a high frequency of false positives and false negatives are the results of their limited adaptability, reliance on manually set rules, and incapacity to manage the complexity of real-world data. This research suggests a novel Deep Attention Network (DAN)-based fraud detection approach to get around these restrictions. The model analyses complicated and high-dimensional healthcare claim data by utilizing sophisticated deep learning approaches, including attention processes. By doing this, it improves the accuracy and dependability of fraud detection by learning to recognize and rank the most pertinent characteristics linked to fraudulent activity. This clever technology can lessen the workload associated with manual intervention and adjust to changing fraud strategies. The ultimate goal of the suggested Deep Attention Network is to offer a precise and scalable solution that improves the security, effectiveness, and transparency of healthcare systems.

7. Bacterial Blight Diseases (BBD) Detection Using Artificial Intelligence: A Systematic Review

   Rahul S. Navale, Nusrat Khan

   Abstract

   Bacterial Blight Disease (BBD) is a significant threat to pomegranate crops, causing severe yield losses and economic setbacks for farmers. Early and accurate detection of BBD is crucial for effective disease management and prevention. Traditional detection methods rely on manual inspection, which is time-consuming, labor-intensive, and often inaccurate. In recent years, Artificial Intelligence (AI)-based approaches, particularly machine learning and deep learning techniques, have emerged as powerful tools for real-time disease detection in agriculture. This systematic review explores the latest advancements in AI-driven BBD detection from pomegranate leaves and fruits, focusing on image processing, computer vision, and sensor-based methodologies. The review synthesizes research findings on datasets, feature extraction techniques, classification algorithms, and performance metrics used for disease identification. Additionally, it highlights the challenges associated with real-time implementation, such as environmental variability, dataset limitations, and computational constraints. Finally, the study discusses future directions, emphasizing the integration of Internet of Things (IoT), edge computing, and hyperspectral imaging for enhanced precision and efficiency. This review serves as a comprehensive resource for researchers and stakeholders aiming to develop robust AI-driven solutions for real-time BBD detection in pomegranate crops.

8. GSM Based Solar Powered Vehicle Accident Detection System

   Sonali Ghodake, V. C. Shinde, Vaibhavi Patil, Sakshi Diwakar, Pranav Phakade, Yogesh Kusumade

   Abstract

   Road accidents are a significant concern worldwide, necessitating prompt detection and response to reduce fatalities. This project introduces a solar-powered accident detection system that leverages GSM technology to provide an efficient and eco-friendly solution. The system is designed to detect collisions through sensors such as an accelerometer and GPS module, which monitor changes in motion and location When an accident is identified, the system sends an alert containing the vehicle’s exact coordinates to predefined emergency contacts using GSM communication. The integration of solar panels as the primary power source ensures the system’s continuous operation, even in remote areas with limited or no access to electricity. This sustainable approach enhances reliability while minimizing the dependence on conventional energy sources. The prototype was tested under various scenarios to validate its functionality, demonstrating quick and accurate accident detection and timely communication of location details. This system offers a cost-effective and scalable approach for improving road safety, especially in regions where immediate medical assistance is often delayed. Potential applications include personal vehicles, commercial fleets, and public transportation systems. Future developments could involve the use of IoT technologies to enable cloud-based monitoring and real-time analytics, as well as machine learning to improve detection accuracy. By combining renewable energy with advanced communication systems, this project not only addresses critical safety concerns but also promotes the adoption of green technologies in modern transportation.

9. Key Determinants of Soil Erosion: Insights Using Machine Learning Algorithm

   Goldi Jarbais, Pon Harshavardhanan

   Abstract

   Soil erosion presents a critical challenge to agricultural sustainability in India, adversely affecting soil fertility, crop productivity, and overall ecosystem stability. This study leverages Machine Learning techniques, specifically Multivariate Regression Analysis, to systematically evaluate and identify the key factors influencing soil erosion. The dataset, sourced from IIT BHU, comprises 15 environmental and land-use variables, including rainfall intensity, soil type, land slope, vegetation cover, land-use practices, and soil texture.

   By analyzing these interrelated factors, the study aims to determine the most influential contributors to soil erosion. The findings highlight that land slope and runoff are the primary drivers of erosion, significantly impacting soil degradation rates. These insights offer valuable guidance for policymakers and agricultural stakeholders in formulating targeted soil conservation strategies to mitigate erosion risks.

   This research underscores the crucial role of data-driven methodologies in environmental analysis and sustainable land management. By integrating statistical modeling with real-world data, the study contributes to the advancement of precision agriculture and erosion control measures, ensuring long-term soil health and agricultural productivity.

10. A LoRa WAN-Enabled Smart Automation Framework for Industrial Applications: Performance and Efficiency

    Bharati Masram, Aarya Sangekar, Yogita Jaiswal, Ankita Godghate, Archana Tiwari, Harsh Manwatkar, Poorvaj Hinge

    Abstract

    In this paper the study of smart automation system remotely controls and monitors industrial appliances using Long Range Wide Area Network (LoRa WAN) technology. The work presents a novel smart automation system, leveraging Long Range Wide Area Network (LoRaWAN) technology to remotely control and monitor industrial appliances. The proposed system integrates a wireless communication framework, sensor nodes, and a user-friendly smartphone application. Utilizing the ESP32 module, the system establishes a reliable communication link between the smartphone and LoRa module, facilitating seamless control and monitoring. The system efficiency is evaluated through three real-world case studies, demonstrating its capability in appliance switching control, fire detection, and ambient temperature and humidity monitoring. Results show an impressive accuracy of 92.33% and operational range of 3–12 km. The system’s adaptable design and low power consumption render it an ideal solution for smart automation applications.

11. Predictive Energy Management System for Hybrid Electric Vehicles with IoT-Driven Sensors

    Yetigadda Shoyab, Jakka Narayana, A. Anilet Bala, Utkarsh Johari

    Abstract

    Traditional energy management in HEVs poses problems such as reactive fault detection, inefficient power sharing, and the lack of online monitoring, hence leading to loss of energy and unexpected failure of the vehicle. Our proposed system integrates MEMS sensors for monitoring vehicle tilt and motion with adaptive speed control based on the terrain and road conditions. It makes use of a CNN-LSTM hybrid model for predictive anomaly detection; hence, the dynamic increase or decrease in motor speed according to the optimization needed in energy consumption aims at efficient power utilization and perfect safety and vehicles. The proposed hardware system fuses the information from five IoT sensors including temperature, vibration, distance, light intensity, and angle in order to monitor the health of the vehicle while minimizing energy consumption. Using the sensor data from the hardware system 4,000 multivariate time-series samples are collected. The collected data are trained with convolutional neural networks for spatial feature extraction and long short-term memory networks for temporal pattern extraction. Real-time prediction is achieved with processing times under 100 ms, and thus the system is suitable for on-road applications. The results show significant improvements over the conventional approaches, with an improvement of 15% accuracy over single-CNN implementations and a 25% reduction in false positive rates. This research advances intelligent transportation systems by presenting a reliable framework for predictive maintenance and energy optimization in hybrid vehicles.

12. Enhanced Content-Based Image Retrieval Using ConvNeXt with Relevance Feedback

    Varkala Satheesh Kumar, Vijaya Chandra Jadala

    Abstract

    Content based image retrieval (CBIR) has been a practical solution to an image retrieval methodology based on image contents. But CBIR systems are generally subject to constraints due to the semantic gap between the low-level image descriptors and human perception. In this paper, to improve the retrieval results, a novel CBIR architecture and a Convolutional Neural Network (CNN), ConvNeXt, which is the leading CNN and relevance feedback (RF) is proposed. A highly performant deep ConvNeXt able to acquire high level image representations is used in order to extract discriminative features by projecting the images to meaningful embeddings that contain the visual characteristics of the image. With a view to the semantic gap, RF is implemented as a modality that the user can interactively annotate the retrieved images in terms of relevance or irrelevance. This feedback is repeatedly modified to drive the evolution of the feature space in such a way that the retrieval system is progressively able to provide satisfactory results in subsequent times. In accordance with the experimental results on test samples from the sample image data sets, the number shows significant performance gain with respect to the traditional CBIR methods, by combining with the RF of the ConvNeXt. By integrating deep learning for feature extraction and users feedbacks for adaption the proposed system can achieve a higher retrieval accuracy and users’ satisfaction. More importantly, this method leads to better and more responsive CBIR systems and handles the inherent problems of these systems as regards to the content of the image.

13. IOT based Prepaid Energy Meter with Stealing Alert

    Anita Shinde, Vaishnavi V. Sardar, Shweta M. Ubale, Sakshi M. Moholkar

    Abstract

    The study focuses on a pre-payment metering system that simplifies the process of reading meters, processing bills, and collecting payments, making it easier for power companies to collect revenue. This system automatically limits electricity use if bills aren’t paid and eliminates the need to stand in line to pay bills. It connects with authorities through a messaging system and doesn’t need a prepaid card. The system not only provides automated meter reading but also helps prevent electricity theft. It monitors energy use and calculates how much power is consumed. The proposed system helps prevent billing mistakes, late payments, and tampering with meters. By using prepaid meters, households can reduce their energy use by 5 to 10%.

14. Diagnosis of Varicose Veins of the Lower Limbs Based on Infrared Thermography

    S. Sivanandam, S. Dhanush, D. Shanley Brighton

    Abstract

    Varicose veins are enlarged, twisted veins that commonly appear on the legs and are caused by weakened or damaged vein walls and valves. Factors like aging, genetics, pregnancy, obesity, and prolonged standing or sitting can increase the risk of developing varicose veins. It leads to abnormal blood pooling and increased skin temperature around affected areas. This study offers an automated diagnostic system for identifying and categorizing varicose veins utilizing MATLAB image processing algorithms and infrared thermography (IR). The suggested method classifies varicose veins into five phases according to intensity and affected area, detects afflicted areas, and analyzes thermal patterns in infrared images. Using predetermined threshold criteria, a region-of-interest (ROI) mask is used to classify the severity of varicose veins. An overall evaluation based on the highest identified stage is included in the tabular presentation of the analysis results, which are graphically represented by bounding boxes of various colors that correlate to severity levels. This automated system offers a non-invasive, economical substitute for conventional techniques while improving the precision and effectiveness of varicose vein diagnostics.

15. Machine Learning Approaches for 5G Coverage Prediction: A Comparative Study of Algorithms and Feature Contributions

    S. Ravi Chandra, Peruri Sri Vedha, Patiwada Madhu Sahithi, Petta Reshma, Purama Manaswi, Shaik Anjum Ara

    Abstract

    As the development of 5G technology is becoming faster and rapid, predicting coverage areas accurately for optimization of network performance and ensuring connectivity has become important. This paper elaborates various machine learning algorithms used in the prediction of 5G coverage using RF Signal Data. For estimating accuracy levels in various models, Target variable, Bandwidth has been considered. Traditional models include Logistic Regression, K-Nearest Neighbors (KNN), Naive Bayes, Random Forest, Support Vector Machine (SVM), LightGBM, AdaBoost, Bayesian Network Classifier, Multi-Layer Perceptron (MLP), Long Short-Term Memory (LSTM), and so on. Such traditional models are evaluated with further advanced techniques such as stacking, voting classifiers, convolutional neural networks, etc. The aim is to identify the most influential features in the 5G coverage prediction. Through comparative analysis, the performance of these models is benchmarked to outline strengths and weaknesses within each approach. Findings from this research suggest that ensemble methods, particularly Stacking and Voting Classifiers, as well as CNN, possess higher prediction accuracy and robustness that would be helpful for 5G network planning and deployment.

16. Seismic Early Warning Framework Utilizing IoT-Integrated Wireless Sensor Networks

    Jeethuri Rakesh, Jathikarthi Bhuvaneshwar, P. Vishnu Kumar, M. V. Dharma Reddy, Shaik Nazeem Basha

    Abstract

    Seismic events pose a significant threat to both human life and infrastructure, often resulting in catastrophic consequences. Traditional earthquake alert systems rely on human intervention, leading to delays when immediate action is needed. This project aims to develop an IoT-based earthquake detection and emergency response system for buildings, offering a faster, automated solution to improve safety. The system uses advanced seismic sensors, microcontrollers, and actuators to detect tremors in real-time. When seismic activity is identified, the system classifies the event and triggers emergency actions, including unlocking exits, automatically opening windows and doors to facilitate evacuation, sounding alarms, and sending alerts to authorities with key information such as event location and magnitude. A mobile app complements the system by providing real time seismic data visualization and historical analysis, helping occupants and emergency services make informed decisions. The system is rigorously tested to ensure it accurately detects significant seismic events while filtering out minor tremors. By automating key functions and offering continuous monitoring, this system aims to reduce response times, enhance occupant safety, and minimize structural damage during earthquakes. This project provides a vital, IoT-powered innovation to improve earthquake preparedness and resilience in at-risk regions.

17. Voice-Enabled AI System for Baby/Toddler Engagement

    A. Parkavi, Uzma Sulthana, G. Brunda, Abdul Wahab, Dharshan Patil, S. Chandrashekhar, V. L. Manoj, H. Ankush

    Abstract

    In today’s fast-paced world where parents struggle to spend quality time with their children, innovative solutions for early childhood development are critical. This paper offers a voice-enabled AI system designed for children (ages 3–6), addressing the challenges busy parents face in engaging their children. Traditional methods often lack individuality and motivation, especially when time is limited. To fill this space, the project proposes robot as mentoring partners for young students. Includes interactive storytelling, educational games, music-friendliness, music, voice recognition and emotional support. This system employs Natural Language Processing (NLP) and Machine Learning (ML) algorithms to tailor interactions based on a child’s age, interests, and developmental stage. It customizes Raspberry Pi modules, microphones, speakers, and other hardware components to streamline audio processing and AI execution. Offering personal interaction, this project helps parents monitor their children’s progress despite short periods. It provides an environment for exploration and growth, laying a strong foundation for lifelong learning.

18. Intelligent Co-operative Malware Detection Using Deep Learning Model

    Morarjee Kolla, Syed Mohammed Ehdeen Ali

    Abstract

    As the frequency of malware attacks rises, affecting numerous users, businesses, and government entities, the importance of malware detection research has intensified. Contemporary malware variants often adopt evasion techniques such as polymorphism and metamorphism to swiftly modify their behaviors and produce numerous iterations. These new threats mainly consist of adaptations of existing malware, which has led to an increased use of machine learning algorithms for detailed analysis. However, these techniques can be time-consuming, requiring significant feature engineering, learning, and representation. To address this research gap, this study examines various deep learning architectures for malware detection, classification, and categorization using a specific dataset. It also mitigates dataset bias in the experimental process by applying distinct dataset splits to train and test the model independently across different timescales. The primary innovation is the development of a novel deep learning framework aimed at establishing an effective model for zero-day malware detection. The Inception-ResNet-v2 model is employed for analyzing malware families. A detailed comparison indicates that this proposed deep learning framework outperforms traditional machine learning algorithms.

19. Advancing Tibetan Text-to-Speech: Challenges and Innovations

    Tenzin Choedon, Shamanth Nagaraju, V. Rekha

    Abstract

    This initiative aims to develop a platform for Tibetan Text-to-Speech (TTS) technology, addressing the significant demand for this technology for the Tibetan language. The main objective of this project is to create a system that is capable of converting text into natural and good quality speech. Through the compilation of Tibetan text-audio datasets, the project meets the increasing demand for technology that preserves oral traditions and allows Tibetans to communicate with other people interested in the language. The process includes the gathering of varied Tibetan text and audio samples, such as news articles, followed by processing of data through cleaning processes and statistical analysis. A benchmark dataset is created to enable the testing of models. The lack of certain resources for Tibetan TTS is addressed by the development of pre-trained machine learning models specific to acoustic modeling, using the adapted FastPitch model for waveform synthesis through the HiFi-GAN vocoder. The existing models were further trained utilizing features particular to Tibetan phonetics and tonalities. The TTS approach is a key strategy for improving digital accessibility for Tibetan speakers and for safeguarding their cultural heritage; it finds applications in media, education, and communication, thus helping to preserve the Tibetan language in the digital era.

20. Fusion of Cryptographic Algorithms and Steganographic Techniques for Enhanced Protection

    Anreddy Shanmukha Reddy, Kolla Sri Ram Charan, Nissankararao Naga Sai, N. Bheshwanth, G. Sreeraag, U. Kumaran

    Abstract

    In the current world, where new and complex threats are formulated daily, there is nowhere that data confidentiality and integrity are more important. The idea of this paper focuses on integrating cryptography and steganography to create a strong security armor against cyber threats. Coupled with this is the difference where cryptography makes data access impossible to understand to any unauthorized personnel while steganography assumes that the message’s very existence is hidden, and instead of looking for the data, one will find, for instance an image or an audio clip. Such a dual-layered strategy improves data protection noticeably. The Advanced Encryption Standard (AES) that is efficient and has an unmatched symmetric key of key size offers genre data confidentiality and is resistant to the brute force attacks. On the other hand, RSA algorithm proves effective in key management because of safe exchange of encryption keys over the insecure channel due to its asymmetrical cryptographic design. Even though it is very powerful, cryptographic techniques alone may stumble when facing attacks on implementations or bad key management. To reduce such risks, this paper will recommend the use of steganography alongside cryptography. Steganography hides encrypted data and thus alarms the potential attackers and provide extra level of protection. With assistance of such methods the approach uses advantages of both cryptography which encrypts the data and steganography which conceals it providing the organism defense from modern threats. This fusion provides a security paradigm that is capable of addressing the new security challenges facing the information society.

21. Symptom Based Medicine Recommendation System

    Abhijeet V. Naik, Shrinidhi Sunadholi, Ujwal M., Kaushik Mallibhat

    Abstract

    The proposed work aims to address prescription discrepancies, a significant challenge in global healthcare, with approximately one in thirty patients affected by errors, as reported by the World Health Organization. To tackle the issue, study presents a comprehensive approach to developing a framework that can predict diseases based on patient symptoms and recommend various medications by employing an ensemble machine learning method. The proposed model is trained on a dataset of diseases, symptoms, and associated medications, and an ensemble method called Bagging (Bootstrap Aggregating) is used, which enhances model stability, reduces variance, and improves accuracy. Bagging also prevents overfitting, especially for high-variance models like the Decision Tree Classifier. The result shows substantial improvements in recommendation compared to conventional single-model approaches like the Decision Tree Classifier which achieved 71.13% accuracy which increases to 98.91% after bagging, demonstrating its efficacy. However, it doesn’t show any improvement in model like Multinomial Naive Bayes which achieved 96.72%, which is already a low-variance model that handles categorical symptom data, boosting model performance on textual or labeled inputs. The model’s predictions are validated using metrics such as F-1 score, precision, and recall. Additionally, a Graphical User Interface (GUI) was developed using the Tkinter library, allowing symptom selection via checkboxes to overcome the errors caused by misspellings or incorrect inputs. The work has implications for personalized healthcare and telemedicine platforms, which will provide a scalable, efficient, and user-friendly tool to assist clinicians and patients.

22. Automated Pomegranate Disease Detection and Classification Using Vision Transformer

    Vidya Shejwal, K. J. Karande, A. C. Pise

    Abstract

    Pomegranate crops are highly susceptible to bacterial, fungal, and viral diseases such as bacterial blight, wilt, anthracnose, and fruit rot, which adversely affect yield and quality. Traditional visual inspection methods are time-consuming and unreliable. To address this, an automated disease detection and classification system is proposed using deep learning and image processing. The system starts with image enhancement using CLAHE, noise removal, and resizing. An Attention U-Net model segments diseased regions from healthy fruit areas. These segments are then classified using a Vision Transformer (ViT), which captures complex features and long-range dependencies for accurate disease recognition. Grad-CAM is used for visualizing the decision-making areas of the model, enhancing transparency. The model is trained on a proprietary dataset containing multiple disease types and healthy samples. It is evaluated using accuracy, precision, recall, F1-score, and IoU. Designed for real-time field use, the model is deployable on mobile devices via Tensor Flow Lite or ONNX.

23. Maritime Obstacle Segmentation for Unmanned Surface Vehicles: A Semantic Approach to Navigating Dynamic Environments

    Ganesh Pawar, Aditya Narthi, D. Shreyanand, Kaushik Mallibhat

    Abstract

    The increasing demand for high-precision navigation in Unmanned Surface Vehicles (USVs) underscores the critical role of image segmentation in maritime environments. Accurate segmentation of maritime images into distinct categories, including sky, water, and obstacles, is essential for ensuring safe operations, mitigating collision risks, and enhancing autonomous decision-making capabilities. To address the challenges, the proposed study presents an advanced image segmentation methodology tailored for maritime applications. The proposed approach leverages a deep learning-based encoder-decoder architecture built upon the ResNet34 framework, incorporating atrous convolution techniques to facilitate multi-scale feature extraction—an essential requirement for effectively capturing the complexity of maritime environments. The architecture further integrates skip connections and upsampling layers within the decoder to enhance segmentation precision and improve predictive accuracy. The implementation framework includes preprocessing techniques such as image resizing, normalization, and standardization to optimize input quality, while post-processing aligns predictions with pre classification categories to enhance interpretability. The model is trained on the LaRS dataset, which provides a diverse and realistic representation of maritime conditions. Experimental evaluations conducted on real-world datasets demonstrate that the proposed model achieves a Mean Intersection over Union (mIoU) of 95.4%, surpassing the performance of existing segmentation approaches. By addressing the limitations of conventional methodologies, the following research contributes to the development of more robust and reliable USV navigation systems capable of operating effectively in diverse and challenging maritime environments.

24. Classification of Brain Tumor Using SqueezeNet and ShuffleNet Architecture

    Arnav Yadav, N. Veni, K. Chirag, Semal Vats

    Abstract

    Brain cancer is a highly dangerous disease because symptoms do not appear until the disease has already spread to other areas of the body. Thus, early detection and classification of abnormal severity is highly important to reduce mortality due to brain cancer. In this work, the classification of brain tumor is performed using advanced computational approaches that analyze digital medical imaging. Specifically, it identifies whether a brain tumor exists and, if so, characterizes the severity by one of four grades such as mild impairment, moderate impairment, very mild impairment, and no impairment. Toward these aims, two of the latest deep learning architectures such as BT-SqueezeNet (Brain Tumor- SqueezeNet) and BT-ShuffleNet (Brain Tumor-ShuffleNet) are proposed in this study to classify the brain tumor. These models are tuned to yield high accuracy and robustness with little computation. SqueezeNet acts as the primary framework because of its fewer computation parameters; however, the performance of the architecture can be improved further by introducing BT-ShuffleNet. The overall classification accuracy of 96.17% is achieved with the proposed BT-ShuffleNet architecture.

25. Integrating Haversine Distance and DBSCAN Clustering for Spatially Enhanced Real Estate Valuation

    B. Hari Ram Deep Reddy, D. Radha, V. S. Kirthika Devi

    Abstract

    Real estate price prediction enables informed decision-making in markets by accurately modeling spatial linkages and market patterns. This study employs DBSCAN clustering with the Haversine formula to calculate geodesic distances and identify clusters within a 10-km radius. A hybrid method computes cluster average pricing, using the median for high-variance clusters and the mean for low-variance clusters. Machine learning models such as Decision Tree, Random Forest, and XGBoost are trained using the generated features and normalized data for real estate price prediction. The best-performing model achieves an accuracy of less than 1% MAPE (Mean Absolute Percentage Error), ensuring reliability. To enhance usability, the proposed system integrates an intuitive user interface for real-time predictions, making it accessible to a broad audience. By combining machine learning, feature engineering, and geographic clustering, this approach significantly improves real estate valuation accuracy.

26. Leveraging BERT and Sentiment Analysis for Enhanced Stock Price Prediction

    B. Krishna Sahithi, D. Radha, V. S. Kirthika Devi

    Abstract

    The nature that is complex and volatile of market has made it difficult to accurately determine the value of stocks. The work present a mixed approach which integrates sentiment analysis with Long Short-Term Memory alongside BERT architecture for stock price prediction purposes. Long Short-Term Memory and BERT architecture in improving stock price prediction methods. By considering quarter earnings call transcripts as the vocalization of market analysts and including investors sentiment subtleties, BERT performs remarkably well in emotion detection. This is combined with historical stock data in an LSTM network to explain temporal behavior patterns in investor sentiments. The model utilized impactful images for predicting future stock prices complemented by explanations that make understanding the forecasted trends easier. The work spotlights the significance of NLP and deep learning techniques in analyzing the financial statements and knowing their significance within the market sectors.

27. Panoptic Segmentation in Unmanned Surface Vehicles

    Spoorti B. Kurubar, G. G. Aiswarya, Sabha K. Gurikar, Kaushik Mallibhat

    Abstract

    The proposed work aims to implementing panoptic segmentation to classify the maritime scene into foreground and background. Such classification is a major challenge for applications like observing the environmental challenges, maritime control, and autonomous navigation to ensure safe and effective operations at various maritime scenes. Utilizing the R-CNN ResNet-101 and Panoptic FPN model on the COCO dataset, the model is trained using the Detectron2 library for object identification and segmentation. It is intended to particularly parse a variety of object types in complex maritime environments, such as obstacles, boats, and ships. The model’s performance in identifying and classifying maritime items has been demonstrated by the training process, achieving an average accuracy of 72%. The result demonstrates De-tectron2’s adaptability in various environmental conditions, which qualifies it for practical maritime applications. The result highlights the effectiveness of Detectron2 in maritime object detection and segmentation, as well as its adaptability in different environmental conditions.

28. Dynamic Multi-attribute Interest Learning for Personalized Product Search

    Khushali Sandhi

    Abstract

    Personalized product search enhances user experiences by tailoring search results to individual preferences based on search logs. Traditional methods emphasize extracting features to build interest profiles but often fail to account for the dynamic variations in user attention to different product attributes (e.g., brand, category). These approaches typically combine all attribute features, relying on models to discern useful patterns from complex scenarios, which limits their effectiveness in capturing nuanced user preferences. To address this gap, we propose a Dynamic Multi-Attribute Interest Learning Model that captures the influence of individual attributes on user interests. Our model introduces two distinct profiling modules: attribute-centered profiling, which identifies preferences for specific attributes, and attribute-aware profiling, which explores multi-attribute correlations within the user’s search history. Additionally, we design a dynamic contribution weights strategy to explicitly guide the model in evaluating the impact of various attributes on user preferences. Experimental evaluations on large-scale datasets demonstrate the superiority of our approach, achieving significant improvements in search accuracy and user satisfaction compared to existing methods. This work advances the understanding of personalized product search by effectively modeling dynamic user interests and attribute interactions.

29. Efficient Motion Regulation of a 6 DOF Arm Utilizing MATLAB

    T. Barathkumar, P. Dayalan, R. Logesh, D. Ruthreshwaran

    Abstract

    This paper explores the framework, modeling, and practical deployment of a six degree-of-freedom (6DOF) system using MATLAB for industrial automation and quality control. The robotic arm is developed to perform precise tasks activities like moving materials, assembling parts, and, enhancing efficiency in automated processes. MATLAB is utilized for kinematic and dynamic analysis, trajectory planning, and real-time simulation, ensuring optimal performance. The control system employs Pulse Width Modulation (PWM) for accurate motor actuation, with coordination managed by an embedded microcontroller. Additionally, an automated quality inspection system is integrated, enabling defect detection and removal with high accuracy. The experimental findings confirm the efficiency of the proposed system in improving automation reliability and operational consistency. This study highlights the role of MATLAB-based removal with robotic solutions in advancing industrial automation and underscores the potential of intelligent robotics in modern manufacturing environments.

30. Enhancing Crisis Response and Misinformation Detection Using Large Language Models and Hybrid Approaches

    Sri Rekha Uppuluri, Sasanko Sekhar Gantayat

    Abstract

    The quick rise of digital communication, especially on social media sites, has had a big effect on study in areas like crisis tracking, finding false information, and analyzing how people feel about things. This study looks at the progress made in Natural Language Processing (NLP), especially the addition of Large Language Models (LLMs) for quick reaction to crises and finding false information. We test how well different deep learning designs work. These include transformer-based models and mixed methods that use rule-based rules. It shows that a mixed model that combines LLMs with heuristic methods greatly enhances classification accuracy, precision, and memory by looking at crisis-related datasets such as CrisisMMD. Our results show that the suggested mix model is more accurate than pure deep learning methods at detecting misinformation and figuring out how people feel about things (94.5%). Even though mixed models make things easier to understand and lower the cost of computing, problems like reducing bias and being able to change in real time are still areas that need more study.

31. Enhancing Sentiment Analysis with Hybrid Transformer-ML Models: A Deep Learning and Explainable AI Approach

    Nikita Gaur, Sridhar chintala

    Abstract

    Sentiment analysis has emerged as an essential component of Natural Language Processing (NLP) for comprehending views in fields such as social media, economics, and healthcare. Nonetheless, deep learning models, despite their great accuracy, often exhibit a deficiency in interpretability, limiting their practical use in critical decision-making contexts. This paper introduces a Hybrid Transformer-ML model that combines BERT-based feature extraction with conventional machine learning classifiers (Random Forest, SVM, and Logistic Regression) to improve accuracy and explainability. The suggested approach guarantees transparency in sentiment categorization by using SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Approach-Agnostic Explanations). Experimental findings across several benchmark datasets indicate that the Hybrid Transformer-ML model surpasses independent deep learning models, with a 95.3% accuracy rate while preserving good interpretability. The results indicate that this method is appropriate for applications necessitating both accuracy and clarity, including financial sentiment analysis and healthcare surveillance.

32. Backmatter