Deep Learning-Based Multimodal System for Behavioral Analysis in Criminal Investigations Shouvik Baidya, Anindya Sundar Layek, Bitan Misra, Swagata Paul, Latika Rahul Desai, Nilanjan Dey International Conference on Emerging Trends and Innovations in ICT Icei 2026, 2026 The field of recognizing human emotions through artificial intelligence (AI) has emerged as a significant and rapidly advancing field of research. In particular, humancomputer interaction (HCI) and affective computing have become the key areas in the field for attempting to sense and understand human affective states and recognizing emotions. Emotional expression is often conveyed through ambiguous, nonverbal, and indirect means, which often require complex modelling and analytical methods. In a multimodal context, emotion recognition integrates and considers multiple sources of signal information and inputs from written text, sounds, vocally, and facial clues to derive human emotional states with even higher accuracy and validity. This research focuses strictly on the approach of recognition of emotions directly, which is based on processing and measuring information from multiple modalities by employing a more advanced, integrated system of emotion analysis called the multimodal emotion and impression analyser (MEIA). Through the combination of multiple data streams (text, audio and video), MEIA is able to develop a detailed profile of a subject's emotional state and general impression. MEIA is publicly available through a gradient-based user interface that enables the user to ingest video by real-time webcam capture or by file upload for realtime and accessible emotion analysis. The multimodal emotion and impression analysis capabilities of MEIA offer promise for use in areas such as criminology, where a better understanding of emotional responses could assist with investigative and forensic processes.
Analyzing Local Texture Patterns for Estimating Tropical Cyclone Intensity Over the North Indian Ocean Pinto Das, Diganta Das, Chinmoy Kar, Swagata Paul, Shiladitya Munshi Proceedings of the IEEE International Conference Image Information Processing, 2025 This study investigates the effectiveness of five local texture descriptors—Local Binary Pattern (LBP), Completed Local Binary Pattern (CLBP), Local Ternary Pattern (LTP), Median Binary Pattern (MBP), and Multi-Modal Local Binary Pattern (MMLBP)—for classifying tropical cyclone intensity using satellite imagery. Each descriptor was evaluated in combination with five machine learning classifiers: Naive Bayes, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree, and Random Forest. The models were assessed using accuracy, F1 score, and ROC AUC metrics. This study shows that the LTP with Random Forest classifier achieved the highest accuracy of 90.71%. The results demonstrate that texture-based features combined with classical machine learning approaches can provide high accuracy, with the added benefits of simplicity and the ability to work on a small dataset, making them suitable for real-time meteorological applications. Future work will explore the integration of time-series data and additional meteorological parameters such as sea surface temperature and wind speed to further enhance prediction performance.
An Exploration of GAN Applications in Stock Prediction through Bibliometric Analysis and Cooccurrence Networks Nabanita Das, Bitan Misra, Bikash Sadhukhan, Pratima Sarkar, Debraj Chatterjee, Swagata Paul 2025 IEEE 14th International Conference on Communication Systems and Network Technologies Csnt 2025, 2025 Predicting the stock market has become a global problem in recent years, requiring massive scientific study. In this study, bibliometric investigation was performed to comprehend a essential outlook and help academics recognize the features of stock market prediction. In light of the aforementioned research, this article concludes with some discussions regarding the upcoming challenges of stock market prediction. With a focus on its many approaches, this article provides a detailed introduction to bibliometric methodology. Additionally, it offers a set of trustworthy, detailed directions for carrying out bibliometric analysis with assurance. Additionally, this study looks into the appropriate application of bibliometric analysis as a substitute for systematic literature reviews. Concretely, it provides a thorough overview of the field of stock market prediction using GAN in terms of the number of regular publications produced, the primary focus of stock market prediction researchers, the most influential nations (authors, institutions, and sources), and intriguing research avenues in this area. The evolution of stock market prediction using GAN is objectively understood in this study, which also serves as a useful reference for academics and researchers in the fields of technology, management, and economics.
Identifying Taxonomy and Assessing Biodiversity from eDNA Datasets Srijita Bhattacharya, Debraj Chatterjee, Sandipan Ghosal, Nabanita Das, Swagata Paul, Bikash Sadhukhan 2025 IEEE Pune Section International Conference Punecon 2025, 2025 The deep ocean hosts vast but poorly characterized biodiversity, where traditional sampling and reference-based eDNA pipelines remain limited. Environmental DNA (eDNA) provides a non-invasive alternative, yet database dependence leads to misclassification, unassigned reads, and underestimated diversity. To address these challenges, we propose an AI-driven pipeline that integrates supervised deep learning and gradient boosting for taxonomic classification, confidence thresholding for reliability, and unsupervised clustering for novel taxa discovery. When applied to the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 8 S}$</tex> rRNA and COI datasets, the system demonstrated high performance: 100% accuracy for Arthropoda and Echinodermata, moderate accuracy for Cnidaria, and lower performance for Porifera due to its sparse representation. Over 95% of the predictions exceeded the 0.90 confidence threshold, ensuring robust classification, whereas low-confidence reads were systematically clustered into novel groups. t-SNE and PCA revealed coherent taxonomic clusters, with some sequences validated as novel or divergent taxa absent from public databases. The pipeline also produced biodiversity profiles that captured community composition and abundance patterns consistent with ecological expectations. By reducing reliance on incomplete databases and improving computational efficiency, the proposed approach enables accurate detection of known taxa and systematic discovery of hidden lineages, offering a scalable framework for biodiversity assessment, conservation, and novel species identification in deep-sea ecosystems.
WhisperSum: Unified Audio-to-Text Summarization Sayam Ganguly, Sourav Mandal, Nabanita Das, Bikash Sadhukhan, Sagarika Sarkar, Swagata Paul International Conference on Intelligent Algorithms for Computational Intelligence Systems Iacis 2024, 2024 In an era overwhelmed by information, efficiently extracting relevant data from audio sources is crucial. This study introduces WhisperSum, a solution combining OpenAI's Whisper for accurate voice transcription with spaCy's advanced natural language processing (NLP) for extractive text summarization. WhisperSum transcribes audio files into text, analyzes the content, and produces concise summaries to help users quickly understand key information. The web application, built using Flask, offers a user-friendly interface for easy audio uploads, real-time transcription, and summary retrieval. The system integrates Whisper's transcription capabilities with the NLP tools of spaCy, ensuring accuracy and efficiency. WhisperSum achieved high performance, with precision, recall, and F1-scores of 91.98%, 92.12%, and 92.04%, respectively, highlighting its effectiveness in accurately transcribing and summarizing audio content while minimizing errors and maintaining comprehensive coverage. The application is valuable for content management, education, and journalism, providing a streamlined solution to information overload. WhisperSum enhances decision-making by ensuring the accessible availability of crucial information, demonstrating the significant potential of integrating advanced NLP and transcription technologies.
A statistical signal processing approach in wireless network traffic analysis Sajib Chowdhury, Swagata Paul, Debraj Chatterjee, Somenath Mukherjee, Sandipan Ghosal, Radha Tamal Goswami 2018 International Conference on Computing Power and Communication Technologies Gucon 2018, 2019 Two vital statistics of wireless network namely peak hour call initiated and call drop have been chosen to examine the self similarity and stationarity behaviour of typical wireless network data in this paper. The scaling pattern and nature of fluctuating frequency are exposed through these two parameters. For exposing the scaling nature of the time series that has been taken from the period 3rd March, 2005 to 31st October, 2015, from the local mobile switching server. Statistical methodologies like Rescaled Analysis (R/S) and General Hurst Estimation (GHE) method are being used to detect the scaling nature of the data-series. Both the time series represent the Short Range Dependency (SRD) and anti-persistency behaviour. The stationarity or non-stationarity behaviour of the time series have been examined by Kwiatkowski Phillips Schmidt Shin (KPSS) test and Continuous Wavelet Transform (CWT). Here both the time series shows non stationarity behaviour.
