Authors - Varalatchoumy M, Syed Hayath, Dinesh D, Dhanush C P, Manu R, V Sadhana Abstract - This paper presents an advanced Generative AI-powered system for video-to-text summarization, leveraging state-of-the-art Computer Vision (CV) technologies and Natural Language Processing (NLP) techniques. The developed system addresses the growing need to extract key information efficiently from lengthy videos across diverse domains such as education, entertainment, sports, and instructional content. By integrating visual and textual data, it pinpoints essential moments and generates concise summaries that capture the core message of the video, reducing the time users spend understanding extensive media. At the heart of this system lies a robust, open-source large language model (LLM), finetuned to produce human-like summaries from video transcripts. The system processes visual cues using advanced CV techniques—such as keyframe extraction and scene segmentation—and textual cues via Automatic Speech Recognition (ASR), which converts audio into text. This dual approach facilitates a deep understanding of spoken and visual content, ensuring that summaries are precise, relevant, and contextually accurate. The system has been evaluated on a diverse dataset, comprising videos of various genres, qualities, and lengths, demonstrating its capability to generalize effectively across a wide spectrum of content. Applications of this video summarization tool include content management, video indexing, educational platforms, and beyond, offering significant time-saving benefits to users and organizations. By incorporating real-time feedback, the system continuously refines its summarization techniques, enhancing accuracy and ensuring that users quickly access the most relevant information, thereby promoting greater accessibility and usability of video content.
Authors - Chetana Shravage, Shubhangi Vairagar, Priya Metri, Shreya C. Jaygude, Pradnya P. Sonawane, Pradnya D. Kudwe, Siddheshwari S. Patil Abstract - This project presents an innovative phishing detection system that addresses the limitations of traditional methods by combining URL-based and content- based features to accurately identify fraudulent websites. Unlike conventional approaches that rely heavily on blacklisting and heuristics, which struggle with zero-day attacks and frequent updates, this system employs machine learning algorithms to automatically extract and analyze critical features from URLs and webpage content. By leveraging a comprehensive dataset that consists of fraudulent (phishing) websites along with legitimate websites, the system aims to improve detection rates to optimize performance based on evaluation metrics like accuracy, precision, F-1 score, recall, and false-positive rates. The system makes use of selective machine learning models like Random Forest, Decision Tree, and Support Vector Machine (SVM), which provide the benefit of increased scalability, robustness and improved effectiveness in phishing detection. Ultimately, this project aims to deliver a scalable, real-time detection solution that effectively mitigates phishing threats in a rapidly evolving landscape.
Authors - Chetana Shravage, Shubhangi Vairagar, Priya Metri, Rohit Rajendra Kalaburgi, Harsh Anil Shah, Abhinandan Vaibhav Sharma, Shubham Shatrughun Godge Abstract - Heart sound categorization is critical for the early identification and detection of cardiovascular illness. Recently deep learning methods have resulted in promising improvements in the correctness of heart sound classification systems. This work introduces a unique transformer-based model for heart sound classification that uses powerful attention mechanism to capture both local as well as global dependencies in heart sound data. Transformers, in contrast to traditional models that rely on handmade features or recurrent networks, can dynamically focus on the most important characteristics in time-series data, making them perfect for dealing with the complexity and variability of phonocardiogram (PCG) signals
Authors - Shubhangi Vairagar, Chetana Shravage, Priya Merti, Nikhil M. Ingale, Sakshi N. Gaikwad, Dhruv G. Yaranalkar, Atharva R. Pimple Abstract - Artificial Intelligence, specifically large language models and generative AI, has dramatically changed the finance industry. According to a few research studies, the current article discusses some of those findings that provided insight into different applications of artificial intelligence in the sector of financial operations and improved predictive analytics, operational efficiency, and quality in decision-making processes. Most critical findings point towards the efficiency of LLMs, where it has achieved automation, financial analysis improvements, and does comply with enforcement standards in their usage of such technologies as towards most security concerns, privacy, and ethical perspectives. Specifically, the long-term implications for financial decision-making and the potential consequences arising from the use of such technologies in an ethical view stand out starkly as red flags of concern. Thus, the review brings new knowledge in the sphere of AI in finance and grounds further justification to be done with proper research motives toward complete responsible development of AI.
Authors - Ashutosh Patil, Gayatri Bhangle, Sejal Kadam, Poonam Vetal, Prajakta Shinde Abstract - Advances in deep learning have enabled effective applications in agriculture, including fruit disease detection. Accurate identification of diseases in fruits such as Annonaceae and Rutaceae families is crucial for yield and quality improvements. Many studies employ deep learning models like CNN, ResNet, VGG, and DenseNet for disease detection across fruits such as apple, orange, guava, and grapes. This article reviews recent research on deep learning for fruit disease detection and classification, focusing on model performance, data utilization, and visualization techniques. We analyze existing studies to identify optimal strategies for fruit species and other underrepresented crops, outlining challenges and areas for future research on various types of fruit species and their family.
Authors - Susheela Vishnoi, Monika Roopak, Prashant Vats Abstract - Pathological tissue image categorization is essential in medical diagnostics, offering insights into disease types, progression, and treatment alternatives. The significant variability in tissue morphology and the overlapping visual patterns across different classes complicate accurate categorization. This study introduces an improved categorization model utilizing a bag-of-features (BoF) methodology integrated with the Roulette Wheel Whale Optimization Algorithm (RWWOA) to enhance classification accuracy and optimize feature selection efficiency. The proposed model utilizes the Bag of Features (BoF) technique to extract discriminative features from tissue images, thereby generating a feature-rich dictionary that represents various pathological structures. The RWWOA is employed to optimize feature selection, thereby reducing dimensionality and concentrating on the most pertinent features for precise categorization. Our method integrates the exploration capabilities of the Whale Optimization Algorithm (WOA) with the probabilistic selection mechanism of the roulette wheel, thereby dynamically balancing exploitation and exploration, which enhances convergence speed and categorization accuracy. Experimental results indicate that the RWWOA-BoF method outperforms traditional methods across various datasets, showing enhancements in classification precision, recall, and F1-score. This method offers a reliable resource for aiding pathologists in diagnostic imaging, which may expedite diagnostic processes and improve consistency in clinical practice.
