Authors - Shailesh Gahane, Payal Khode, Arya Kapse, Deepak Sharma, Pankajkumar Anawade Abstract - The global health care is on the threshold of a revolutionary transformation, and the artificial intelligence technology stands at the forefront of this change. This research paper deals with the complex process involved in conceptualizing, designing, developing, and enforcing an AI-pushed health-care system. With the strength of machine learning and deep learning technologies, it can analyze vast ranges of healthcare data, which incorporates digital fitness records, clinical imaging, among others. It begins with reviewing multidimensional literature pertinent to the research study. Through this research, a critical part entails a pilot observe designed in a very deliberate manner in order to conservatively test the effectiveness and reliability of AI algorithms in various healthcare fields. Based on this research, it is predicted that several advantages are bound to be realized among them; more accurate diagnosis, individually tailored treatment plans, optimized effective care resources deployment, and lower healthcare expenses. By making known my findings and insights, I hope to provide helpful guidance and recommendations for health care professionals, policymakers, and developers of technology, which would eventually enrich or enhance the discourse regarding AI integration in health care.
Authors - Ankit Shah, Hardik M. Patel Abstract - Generative Adversarial Networks (GANs) have revolutionized data augmentation by generating realistic and diverse synthetic data, significantly enhancing the performance of machine learning models. This review evaluates the efficacy of GAN-based augmentation compared to traditional methods across various datasets, including MNIST, CIFAR-10, and diabetic retinopathy images. Using architectures such as DCGAN, WGAN-GP, and StyleGAN, our experiments showed substantial performance improvements: CNN accuracy on CIFAR-10 increased from 82.0% to 87.5%, and ResNet-50 accuracy on diabetic retinopathy images rose from 75.0% to 87.0%. Statistical analyses confirmed the significance of these gains. Despite challenges like computational costs and training instability, GAN-based augmentation proves superior in addressing data scarcity and enhancing model robustness. Future research should focus on optimizing GAN training, integrating hybrid models, and exploring ethical considerations. The results underscore GANs' potential in advancing machine learning applications, particularly in complex and data-scarce domains.
Authors - Manoj N, M Thanmay Ram, Manikanta S, Tarun Pradeep, Ramandeep Kaur Abstract - As IoT devices multiply in smart cities, safeguarding healthcare data's confidentiality, security, and integrity from various sources is getting harder. In order to protect healthcare data and facilitate effective machine learning, this article suggests a secure structure that combines Blockchain technology with Federated Learning (FL). With its immutable ledger, blockchain guarantees data confidentiality and openness throughout the network, whereas FL lets data stay on local devices, protecting privacy while training models. The suggested framework is ideal for smart city applications since it places a strong emphasis on safe data sharing, privacy protection, and dependable model management. The design tackles important problems like data breaches, illegal access, and confidence in model updates by utilizing FL's decentralized training and Blockchain's tamper-proof data management. This combination promotes openness and confidence among stakeholders while strengthening the security of healthcare data. The suggested method, which is intended for smart cities, opens the door for creative and privacy-compliant approaches to healthcare data administration and analysis by facilitating efficient collaboration across healthcare organizations without compromising patient privacy.
Authors - Sabina Sehajpal, Ravneet Kaur, Ajay Singh, Mukul Bhatnagar Abstract - This research elucidates the transformative potential of big data analytics and artificial intelligence in optimising health insurance claims and risk assessment by employing an empirically robust framework encompassing reliability and validity metrics, Heterotrait-Monotrait Ratio (HTMT) analysis, and bootstrapping to unravel the intricate interdependencies among constructs such as AI model accuracy, claims processing efficiency, cost efficiency, data quality, fraud detection accuracy, system usability, and user trust interface, thereby advancing a comprehensive understanding of the systemic synergies that enhance predictive precision, operational scalability, and equitable resource allocation within the healthcare financing paradigm.
Authors - Prajakta Deshpande, Divya Kasat, Shrushti Mahadik, Rutika Ubalekar Abstract - Mantras in the Sanskrit language are the soul of Indian culture, carrying deep spiritual, emotional, and cultural implications. These ancient chants, more than words, resonate profoundly and are used in meditation, healing, and divine invocation. Each Sanskrit mantra reveals emotional connotations through its meaning and sound. We have classified them into three groups: Vidur Niti, representing clarity and wisdom; Chanakya Niti, embodying planning and decisive action; and Sanskrit Shlokas, symbolizing harmony and unity. In pioneering work, cutting-edge transformer models, such as XLNet, and the Hugging Face framework are adapted to build an advanced text classification system that decodes the emotional essence of sacred mantras. A hand-curated dataset of annotated Sanskrit mantras has its performance evaluated in terms of accuracy and F1-score on emotional polarity. This kind of research bridges ancient wisdom with modern technology to uncover the revitalization of sacred traditions through computational linguistics for this very modern world.
Authors - Panchal Twinkle Shaileshbhai, Pushpal Desai Abstract - Shadow rendering plays a crucial role in enhancing the realism and immersion of Augmented Reality (AR) applications by seamlessly integrating virtual objects into real-world environments. Dynamic daylight conditions, characterized by varying sunlight intensity, direction, and ambient light, present significant challenges to achieving visually coherent and computationally efficient shadow rendering. This study offers a comparative analysis of diverse shadow rendering mechanisms, evaluating their effectiveness, performance, and suitability for AR applications under fluctuating lighting conditions. Techniques such as Light Direction Approximation, Shadow Mapping, Projected Planar Shadows, and Real-Time Ray Tracing, Dynamic Shadow Blending, Real-Time Sun Position and Shadow Adjustment, Hybrid Shadow Techniques, Brightness Induction and Shadow Inducers and Shadow Perception in AR are examined, highlighting their strengths, limitations, and application scenarios. The research also addresses factors influencing shadow intensity and alignment, providing insights into optimizing realism and computational efficiency in outdoor AR environments. By exploring innovative solution and proposing guidelines for shadow rendering mechanism, this study contributes to advancing AR technology, ensuring enhanced visual fidelity and user experience across dynamic settings.
Authors - T. A. Alka, M. Suresh, Aswathy Sreenivasan Abstract - This study aims to explore entrepreneurship trends in computer science through Bibliometric analysis. 5530 documents from the Scopus are selected based on inclusion and exclusion criteria in the initially selected documents. The Biblioshiny package under R programming is used for the analysis. The major findings are; entrepreneurship has wider applications in various domains. It is not a single-domain phenomenon. The trend topics and word cloud show the most trends in entrepreneurship in computer science including learning models, artificial intelligence, games, innovation, entrepreneurship education, digital transformation, computer simulation, etc. The limitations of the study are; papers from the Scopus database are only considered. Documents other than in English, and papers from other domains except computer science are ignored. This literature study lacks the benefits of primary data research. The inherent limitations of the bibliometric methodology will affect the results. The findings of this research provide knowledge on various aspects to the policymakers, practitioners, researchers, and academicians to foster an entrepreneurship ecosystem and understand the trends of entrepreneurship in the computer science domain. The novelty of the study is underlying the comprehensive review of the existing body of knowledge to draw future research directions. The main highlight of this literature review paper is that complete in-depth knowledge of the data is possible through bibliometric analysis.
Authors - Sathiyapriya K, S Bharath, Rohith Sundharamurthy, Prithivi Raaj K, Rakesh Kumar S, Rakkul Pravesh M, N Arun Eshwer Abstract - The convenience and security offered by voice-based authentication systems results in its increasing use in various sectors such as banking, e-commerce, telecommunications, etc. But these systems are open to vulnerabilities from voice spoofing attacks, including replay synthesis and voice conversion. The following work makes use of Mel-Frequency Cepstral Coefficients (MFCC), Constant-Q Transform (CQT), and a deep learning model Res2Net and creates a framework that can classify genuine and spoofed voices. MFCC and CQT are commonly used for feature extraction, and the Res2Net model classifies the audio. The system was evaluated against the ASVspoof 2021 dataset, the reason being that it has a diverse collection of audio samples (almost 180,000) samples, and also it is recognized by the research community. Our system recorded a low Equal Error Rate (EER) of 0.0332 and a Tandem Detection Cost Function (t-DCF) of 0.2246. This framework contributes to the advancement of secure voice authentication systems, addressing critical challenges in modern cybersecurity.
Authors - Martin Mollay, Deepak Sharma, Pankajkumar Anawade, Chetan Parlikar Abstract - This study examines how AI affects consumer choices in smart homes. This research determines how AI-supported technologies such as voice-controlled digital assistants, dynamic pricing models, and personalized recommendations significantly affect consumer tastes, behaviors, and purchasing decisions through the use of secondary data sources. Customers’ interactions with goods and services are personal and pragmatic as artificial intelligence is progressively included in smart homes. The study claims, however, that artificial intelligence has a two-edged effect on consumer decision-making. Two such areas where AI can enhance customer experience by improving interactions and decision-making processes are through personalization and optimization. This, however, gives rise to some critical ethical issues concerning algorithmic bias privacy and data security. As technology matures, it is essential to promote responsible AI practices, given its increasing ubiquity in daily life. According to the study findings, for instance, organizations must overcome these challenges if they are to preserve customer trust and ensure that artificial intelligence (AI) will ultimately enhance customer relationships. Reading through many kinds of research, company reports, and scholarly works on AI applications in consumer decision-making gives one a view of its current and potential future applications. Results underline that ethics matter when designing transparent AI systems in order to enhance customer loyalty and trust.
Authors - Payal Khode, Shailesh Gahane, Arya Kapse, Pankajkumar Anawade, Deepak Sharma Abstract - The proposed identity card processing system revolutionizes the traditional, manual, and semi-automated ID card creation processes by integrating advanced web technologies and artificial intelligence (AI). Designed for efficiency and user-friendliness, this system employs React JS and JavaScript for seamless operation, enabling students to input required details and generate a printable ID card within 15 minutes. This contrasts significantly with the time-consuming manual design methods using applications like CorelDRAW or Photoshop. Incorporating AI-driven features such as customizable designs and face detection technology ensures quick and accurate retrieval of student data from the school database. The system emphasizes real-time data processing, cross-platform accessibility, and a secure, intuitive interface, allowing users and administrators to handle ID card requests efficiently from any internet-enabled device. By addressing the limitations of existing methods, this automated solution ensures flexibility, reliability, and enhanced usability, making ID card issuance streamlined and error-free. The final system aligns with modern technical and operational requirements, delivering robust functionality and improved organizational efficiency.
Authors - Mumthaz Beegum M, Raseena Beevi, Aji S Abstract - Short-text similarity is a vital research area in NLP with significant implications for use cases like search recommendations and question-answer systems. Traditional models often focus solely on semantic similarity, overlooking syntactic factors. Our approach uses the AnglE (Angle Embedding) method for semantic similarity, which transforms text into high-dimensional vectors to capture the nuanced meanings and relationships between words and phrases. The cosine similarity measure is then employed to calculate the similarity score from these vectors. We apply the weighted Tree Edit Distance (TED) method for syntactic similarity, which measures structural differences between parse trees by calculating the minimum cost required to convert one tree into another through a series of edit operations. By integrating these two complementary similarity measures, our approach aims to deliver a more thorough and accurate evaluation of text similarity. This methodology introduces an advanced technique that combines semantic and structural information to enhance the assessment of short-text similarity. The integrated methodology introduces a sophisticated framework that not only enhances the precision of similarity evaluations but also bridges the gap between semantic and syntactic analyses, thereby offering a more comprehensive evaluation of text similarity.
Authors - Suhail Manzoor, Rahul Gupta, Prakhar Sharma, Yash Mittal, Mohammad Arshad Iqbal Abstract - Plants are mainly suffering from abiotic stress such as drought, salinity, and widely temperature decrease or increase. Thanks to notable advancements in machine learning and hyperspectral imaging, detecting stress in plants has never been easier. For that matter, different machine learning techniques such as Random Forest, Support Vector Machines (SVM), Convolutional Neural Networks (CNN), and Kernel Ridge Regression have been used. Hyperspectral image has been widely used in classifying crop water stress by classifiers such as Random Forest and SVM. CNNs have been widely used for plant phenotyping under multiple stresses thanks to their good prediction results, but that entails many computation problems. Other methods, such as Kernel Ridge Regression and Extreme Gradient Boosting have emerged to target specific stress indicators like leaf reflectance spectra at key wavelengths; however, these typically rely on specialized equipment and significant data preprocessing. Here, we synthesize these various approaches to plant stress detection and present an integrated approach for abiotic stress recognition in plants and advocate for models with high generalizability across different environmental conditions or types of biotic stresses.
Authors - Shriya Vadavalli, Geethika Bodagala, T Sridevi Abstract - This paper conducts a systematic literature review about the current status of the art in the detection and diagnosis of autoimmune skin diseases. This paper identifies recent studies and advancements in terms of key technologies, methodologies, and approaches used in this domain specifically with regard to imaging and deep learning techniques. Therefore, the work underlines scopes for further studies in terms of improving diagnostics accuracy and increasing robustness and accessibility for diagnostic solutions. This piece of work also puts a hole in the existing literature, and research gaps on machine learning algorithms and image processing have highlighted the enhancement of the precision as well as effectiveness of such detection systems with respect to skin diseases. This review is in a position to provide grounds for future research directions and innovations in skin disease diagnostics.
Authors - Ch. G.M.A.S. Teja, V. Manmohan, Ch. Srinith, N. Shiva Kumar, Vempaty Prashanthi, R Govardhan reddy Abstract - This review paper focuses on the emerging potential of FKP recognition as a strong modality for identity verification and authentication. Traditional biometric methods are fingerprint and iris recognition methods, which have been adopted more than others because they can be accurate and reliable, but these techniques also bear limitations, such as problem with false negatives, costly equipment, and vulnerability to data breaches. In For the response to these challenges, FKP offers a new approach in that the ability to search for unique patterns of knuckle skin is the stable and more non-invasive indicator. Unlike, which can deteriorate and is vulnerable to influences over time, FKP has little change from outer conditions and, therefore, is an attractive solution for secure and contactless authentication. This review synthesizes the most recent research and technical advancements in the recognition of FKP. Presenting the benefit of how bringing FKP within the multimodal system compiles diverse strengths of various biometric techniques under one framework and provides the enhancement of a holistic method toward security. The discussion continues with research gaps among the existing literature and ends with a call for further investigation. All the above said, the review given is going to validate the potentiality of FKP becoming a feasible, scalable surrogate to the traditional biometric schemes in various datasets, real word applications.
Authors - Kiruthika R, Gunavathi N Abstract - This study focuses on a conventional inset-fed rectangular patch antenna to investigate various substrate materials for terahertz (THz) frequency applications. The performance of different substrates is evaluated based on their electrical parameters. The operating frequency range by IEEE standards falls within the THz band, specifically from 0.1 to 3, with a center frequency of 1.5 THz. Key performance metrics such as return loss (in dB), bandwidth, gain, directivity, and efficiency are assessed, along with bending tolerance to evaluate mechanical stability. Teflon demonstrates superior radiation characteristics, achieving high gain and directivity values obtained through high-frequency structural simulator (HFSS) and computer simulation technology (CST). Additionally, based on statistical analysis, Arlon Diclad 880 provides better mechanical stability than other substrate materials. The equivalent circuit model is analyzed using advanced design system (ADS) software.
Authors - Agha Imran Husain, Sachin Lakra Abstract - In this paper, we proposed a novel swarm-based algorithm called the Flying Duck formation Energy Optimization Algorithm (FDEOA) for selecting cluster heads in Wireless Sensor Networks (WSNs). The FDEOA method tries to minimize energy usage in WSNs by lowering the number of messages delivered by the sensor nodes. It is motivated by the formation of a flock of flying ducks. The best sensor node for the cluster head is chosen by the algorithm using a multi-objective fitness function that takes into account both energy usage and network connection. In terms of network longevity, energy usage, and the number of dead nodes, the FDEOA method is contrasted with other well-known clustering algorithms, including LEACH and SEP. Simulation findings reveal that the FDEOA algorithm beats the existing methods in terms of network lifetime and energy usage while retaining a high level of network connectedness. The suggested approach can be used on low-power sensor nodes and is computationally effective. The FDEOA algorithm has the potential to be used in a variety of WSN applications where network longevity and energy usage are important factors.
Authors - K. L. Sudha, Kavita Guddad Abstract - NavIC (Navigation with Indian Constellation) is a satellite system consisting of seven satellites orbiting the Earth in GEO and GSO orbits. This satellite constellation offers Standard Positioning Service (SPS) for common public use and Restricted Service (RS) for approved users, using two frequencies: the L5 band and the S band, with the CDMA technique. This paper examines the suitability of three binary sequences — Gold, Weil, and Weil-like Sidelnikov-Lempel-Cohn-Eastman (WSLCE) sequences — as PRN codes for the primary in phase and quadrature-phase codes of the L5 band of IRNSS. It describes the generation of these sequences and compares them based on their even auto-correlation and cross-correlation values. The randomness of these sequences is evaluated using the NIST (National Institute of Standards and Technology) statistical test suite. A comparison of the three binary PRN sequences, each 10230 bits in length for the L5 frequency band, reveals that WSLCE sequences exhibit greater randomness compared to the other two sequences
Authors - Anupama Nayak, Shikha, Jahanvi Ojha, Kavita Sharma, S.R.N Reddy Abstract - Living in a world where science is advanced, the presence of technology can be seen in even the smallest of the appliances we use and it has become a crucial part of our lives. Technologies like the Internet of Things (IoT) along with modern wireless systems have led to the creation of smart appliances which have made the lives of people much simpler and more comfortable. Many automation devices in our homes are accessible remotely via a mobile application. This paper focuses on the design and development of an android mobile application, and its connectivity to a cloud database, which is also accessed by the hardware. Using Android Studio, Firebase cloud database, and Raspberry Pi, we successfully developed an android application that controls hardware remotely.
Authors - Vaishnavi Moorthy, Rupen, Dhruv Chopra, Anamika Jain Abstract - The issue of security is paramount in any organization. Since the advancement of technology and introduction of Generative AI such as ChatGPT, security concerns have skyrocketed for everyone alike. With the availability of these Generative AI platforms to virtually anyone with internet access the threat of security is bigger than ever before as these platforms can be used for malicious intents by a large number of people. Limited research has been performed by third party researchers on Generative AI as it is a relatively new technology. We intend to perform relative research in identifying potential vulnerabilities in Generative AI models, LLMs etc. The aim of this research is to document various ways cybersecurity can be used in GenAI with the intention of both securing assets and protection against malicious activity. The project also delves into potential applications of GenAI in helping identify, prevent and test various security infrastructure. Some potential threats and uses studied under this project include real time access management, phishing detection, jailbreak of ChatGPT. The mentioned use cases provide us with a wide picture of uses of these LLMs in the world of cybersecurity. The deployment of Generative AI systems introduces significant cybersecurity challenges, necessitating the need of safeguards for monitoring against threats and vulnerabilities. This project aims at performing the necessary research to identify such situations from affecting normal operations of an organization and to spread awareness regarding the use of Generative AI in cybersecurity.
Authors - Aatm Prakash Rai, Puneet Kumar Gupta, Santanu Roy Abstract - The proliferation of Generative Artificial Intelligence Chatbots, also known as Gen AI chatbots, are nowadays in the growth phase of integration with information systems for effective teaching and learning. The learning experience has been enhanced with the arrival of Gen AI tools such as machine learning and natural language processing. Gen AI Chatbots like ChatGPT, Copilot, etc. can be considered as computer programs that can trigger human-like interactions to aid investigating, developing and transferring knowledge. The objective of this research work is to scan the previous scholarly research works on Gen AI chatbots adoption by relying on bibliometric analysis. The study intends to contribute by identifying research trends and insights towards adoption of Gen AI chatbots in higher education sector in the Indian context. The outcome of the analysis highlights prospective research opportunity in Gen AI chatbots due to the evolution of the large language models and machine learning. This emerging technology may alter the course of future research in the higher education sector.
Authors - Amani A. Aladeemy, Sachin N. Deshmukh Abstract - Sentiment analysis (SA) discerns the subjective tone within text, categorising it as positive, neutral, or negative. Arabic Sentiment Analysis (ASA) has distinct obstacles owing to the language's intricate morphology, many dialects, and elaborate linguistic frameworks. This study compares SA models for Arabic text across multiple datasets, evaluating traditional machine learning (ML) algorithms, such as Random Forest (RF) and Support Vector Machine (SVM); deep learning (DL) models, including Bidirectional Long Short-Term Memory (BiLSTM) and Bidirectional Gated Recurrent Unit (BiGRU); and transformer-based models like BERT, AraBERT, and XLM-RoBERTa. Experiments on datasets—HARD, Khooli, AJGT, and Ar-Tweet—covering MSA and dialects such as Gulf and Egyptian demonstrate that transformer-based models, particularly AraBERT v02, achieve the highest accuracy of 93.9% on the HARD dataset. The study highlights the significance of dataset characteristics and the advantages of advanced models, offering valuable insights into Arabic NLP and advancing SA research.
Authors - Ritika Upadhyay, Eshita Dey, Munmun Patra, Roji Khatun, Chinmoy Kar, Somenath Chaterjee Abstract - Predicting accurate rainfall is crucial for a country like India, which has a diverse economy. Agriculture is a vital aspect of life for many rural communities in India, making timely rainfall a significant concern for improving agricultural yields. However, predicting rainfall has become increasingly challenging due to drastic climate changes, resulting in more frequent natural calamities like floods and soil erosion. To address this issue, extensive research is underway to enhance rainfall prediction, allowing people to take appropriate precautions to protect their crops. Currently, predictive models tend to be complex statistical frameworks that can be expensive in terms of both computation and budget. As a more effective solution, using historical data combined with machine learning algorithms is being proposed. This research aims to improve rainfall prediction through algorithms such as Gradient Boosting and Random Forest. Model evaluation will utilize metrics like Mean Squared Error (MSE) and Root Mean Squared Error (RMSE). This study has considered approximately 150 years of historical rainfall data (from 1813 to 2006) for different regions of India.
Authors - G.KALANANDHINI, VIGNESHWARAN.D, R.KARTHIKA, S.PUSHPALATHA, D.SAKTHIPRIYA Abstract - Activity delays confronted by crisis vehicles regularly result in basic time misfortune, imperiling lives. The Help to begin with Responders (AFR) framework addresses this issue by utilizing LoRa SX1276 communication modules designed as transmitters in crisis vehicles to communicate with recipients at activity intersections. This framework empowers programmed green light signals for drawing closer crisis vehicles, guaranteeing continuous section. GPS NEO-6M modules give directional data, whereas a centralized authorization component anticipates abuse. Particular vehicle IDs permit for prioritized reaction, with fire motors taking the most noteworthy need, taken after by ambulances and police cars. Activity policemen are informed to oversee synchronous mediations successfully. Typical activity flag operations continue when no crisis vehicle is recognized. The AFR framework leverages Arduino Nano for LoRa modules, Arduino UNO for activity control, and ESP8266 for authorization. This integration improves crisis reaction times and moves forward security for both patients and responders, displaying a noteworthy progression in urban activity management.
Authors - Varun M V, Venkat Raghavendra A H, V Hemanth, Ashwini Bhat Abstract - The work undertaken is a comprehensive analysis of cricket sounds, focusing on the interaction of the ball with the bat and the wicket, the study aims to distinguish between edged, shot, and bowled audio in both noisy and noise-free environments. Upon feature extraction, machine learning models XGBoost and Random Forest were trained, to accurately classify these distinct cricketing events. This not only enriches the realm of cricket analysis by facilitating informed decision-making and insights into player performance but also showcases the potential of audio-based sports analytics.
Authors - N.Janani Abstract - In clinical practices, almost 18-20% cases go either unnoticed or misdiagnosed due to overlapping and subtle features in imaging, especially in complicated cases. We tackle this by using Cross-Modality Attention Network (CMAN) which integrates details from multi-phase CTs. By leveraging the distinct advantages of each scan phase, this method provides a comprehensive understanding of the tumor’s structure and characteristics. The cross-modality approach employs an attention mechanism to integrate information from multiple scan modalities, each capturing unique details. This process emphasizes the most critical tumor-related features while effectively minimizing noise, ensuring enhanced classification accuracy. Achieving an impressive accuracy of 98.47% on the LIDC-IDRI dataset, the CMAN significantly reduces misdiagnosis in complex cases. This approach can be really helpful in filling the diagnostic gaps, facilitating more informed clinical decision-making and improved patient outcomes.
Authors - Parekh Rikita Dhaval, Hiteishi M. Diwanji Abstract - The Visual Question answering is an emerging multidisciplinary research field that intersect computer vision and natural language processing. Medical Visual Question Answering is one of the prominent area of VQA. Medical images and Clinical Questions related to the medical image is given as input to the VQA model and VQA model respond with corresponding answer in natural language. The aim of Medical VQA is to enhance interpretability of medical image data for enhancing diagnostic accuracy, clinical decision making and patient care. This paper presents a novel framework that integrates Vision Transformer (ViT), Language transformer (BERT), and a Convolutional Autoencoder (CAE) to improve the performance of Medical VQA task. The Vision Transformer is used to capture complex visual features from medical images, while BERT processes the corresponding clinical question to understand its context and generate meaningful language embedding. To further enhance visual feature extraction, a Convolutional Autoencoder [1], [2] is incorporated to preprocess and denoise the medical images, capturing essential patterns, compressing medical images without losing key features, thereby providing cleaner input to the ViT. The combined use of these three components enables the model to effectively align visual features with textual information, leading to more precise and context-aware answers. We evaluate the proposed ViT+BERT+CAE model on benchmark medical VQA dataset MEDVQA-2019, showing significant improvements over traditional methods based solely on convolutional or recurrent networks. The results demonstrate significant increase in accuracy, precision, recall, F1-Score and WuPS score after applying Covolutional AutoEncoder in Preprocessing stage.
Authors - Ritu Raj Pradhan, Darshan Gera, P. Sunil Kumar Abstract - This paper explores static facial expression recognition (FER) and presents a novel facial augmentation technique designed to enhance model training. By utilizing pre-trained facial landmark detection models, we analyze the spatial structure of faces within the FER training dataset. Based on the predicted landmark coordinates, facial images are augmented by strategically masking patches of varying sizes at key landmark locations. This approach emphasizes the structural significance of facial landmarks while preserving other critical facial features, enabling models to capture both global facial structure and nuanced expression-related details. Extensive experiments on benchmark datasets validate the effectiveness of the proposed method, showcasing its potential to improve FER performance, particularly in challenging scenarios.
Authors - Deepali, Karuna Kadian, Kashish Arora, Saumya Johar, Liza Abstract - The stock market has become increasingly unpredictable in recent years due to various factors like public sentiments, economy and geopolitical issues. The Traditional methods being used like time series model and Long Short-Term Memory (LSTM) models, often don’t make the correct predictions as they rely mostly on historical data of stock market and so they fail to grasp how market behaves or how chaotic behavior of market can be analyzed. These models hence may fail in case of making wise investment decisions. Our proposed methodology comes up with a hybrid approach using chaos theory, sentimental analysis for overcoming these challenges by analyzing the how stock prices might change according to the sentiments of people. We analyze 65,000 tweets of 95 organizations and their stocks and use chaos theory to find hidden patterns in stock movements. The classical computers take high computational time to analyze complex problems like stock market predictions. Hence, we combine these approaches with the Quantum Approximate Optimization Algorithm (QAOA) to solve the complex patterns of stock price prediction faster and more accurately than classical methods. We have used sentimental analysis, chaos theory with QAOA which is a combinatorial algorithm, being used to optimize the stock portfolio based on specific stock metrics- inclusive of F1 score(from sentimental anaylsis) and chaos theory assessments, it researches for the organisations with stability and low risk-high returns in stock market. Thus aiding investors and traders to make an informed decisions regarding where to invest with low risk and high returns.
Authors - Dipti Varpe, Kalyani Kulkarni, Vaidehi Deshpande, Vedaant Deshpande, Vaishnavi Habbu Abstract - Augmented Reality (AR) and Virtual Reality (VR) enhance traditional pedagogical methods by providing immersive, interactive and experiential learning environments, while catering to diverse learning styles. The paper examines their effectiveness in improving knowledge retention, fostering engagement, and enabling hands-on practice in simulated real-world scenarios, citing comparisons with traditional teaching tools. In education, AR and VR allow visualization of abstract concepts, collaborative virtual environments and gamified learning experiences that make complex subjects accessible and engaging. For training purposes, these technologies are instrumental in safe skill acquisition, particularly in high-risk fields such as healthcare and military operations. Challenges such as high costs of facility maintenance and safe implementation are also addressed. This review concludes with recommendations for leveraging this technology to create impactful and scalable solutions for learners and trainees in various disciplines.
Authors - R.V. Sai Sriram, A. Srujan, K. Rahul, K. Sathvik, Para Upendar Abstract - Freshness plays a crucial role in determining the quality of fruits and vegetables, directly impacting consumer health and influencing nutritional value. Fresh produce used in food processing industries must go through multiple stages—harvesting, sorting, classification, grading, and more—before reaching the customer. This paper introduces an organized and precise approach for classifying and detecting the freshness of fruits and vegetables. Leveraging advanced deep learning models, particularly convolutional neural networks (CNNs), this method analyzes images of produce. The training and evaluation dataset is large and varied, including diverse fruits and vegetables in various conditions. Freshness is determined by analyzing key features like color, texture, shape, and size. For example, fresh produce typically shows vibrant color and is free from mold or brown spots. Traditional methods for assessing quality through manual inspection and sorting are often slow and error prone. Automated detection techniques can significantly mitigate these challenges. Therefore, this paper proposes an automated approach to freshness detection, which first identifies whether an image shows a fruit or vegetable and then classifies it as either fresh or rotten. The ResNet18 deep learning model is employed for this identification and classification task. It also estimates the size of the fruit/vegetable using OpenCV. The qualitative analysis of this approach demonstrates outstanding performance on the fruits and vegetables dataset.
