Research
Published machine learning research
Unleashing the Power of Dynamic Mode Decomposition and Deep Learning for Rainfall Prediction in North-East India
Temporal and spatio-temporal analysis (2024)
Accurate rainfall forecasting is crucial for effective disaster preparedness and mitigation, especially in the North-East region of India, which is prone to extreme weather events. This project explores the use of Dynamic Mode Decomposition (DMD) and Long Short-Term Memory (LSTM) models to enhance rainfall prediction using 122 years of historical data from the India Meteorological Department. Our study demonstrates that while both methods are effective, LSTM outperforms DMD in capturing complex nonlinear relationships, making it a powerful tool for rainfall forecasting.
The results are officially published at ICCAIML 2024.
Comparative Analysis on Speech-Driven Gesture Generation
Gesture generation for robots based on speech (2023)
Advancing human-agent interaction requires effective gesture generation from speech. This study explores deep learning methods such as Gated Recurrent Unit (GRU), Bidirectional-GRU, and Multi-Head Attention for translating speech into human-like gestures. Using Mel-Frequency Cepstral Coefficients (MFCCs) as input features, our comparative analysis highlights the superior performance of the Bi-GRU model in generating natural gestures. These findings contribute to the development of more intuitive and responsive virtual agents and robots.
The results are officially published at AKIP 2023.
Face Mask Detection using Transfer Learning and TensorRT Optimization
Computer Vision & Embedded AI (2023)
Deploying real-time face mask detection on embedded systems requires efficient inference optimization. This study explores the use of TensorRT to accelerate deep learning models for face mask detection on resource-constrained devices like the NVIDIA Jetson Nano. Using transfer learning with architectures such as MobileNetV3, ResNet50V2, and EfficientNetB0, we demonstrate significant inference speed improvements while maintaining high accuracy. Our results show that the MobileNetV3 model achieves 85 FPS, outperforming previous implementations.
The results are officially published in ICICC 2023.