Nikhil Thapa
Machine learning researcher with three first-author publications in computer vision and signal processing. Work focuses on representation learning and sequence modelling for perception and generation tasks, spanning deep metric learning, Transformers, TCNs, diffusion models, and VQ-VAE. Experience covers the full research cycle from problem formulation through experimental design to publication. Interested in learning-based systems that produce robust, structured representations under real-world constraints.
Research Assistant at AILab, Jeonbuk National University, led by Prof. Dr. Lee Joon Whoan.
Research Interests
Computer Vision · Deep Metric Learning · Generative Modelling · Representation Learning · Sequence Modelling · Applied Deep Learning
Work
Visual Perception under Weak Supervision — Pine wilt disease detection from drone orthomaps using YOLOv8 segmentation followed by deep metric learning (ResNet-50, semi-hard triplet loss) for species classification. Achieves 83% mIoU on segmentation and 98.7% classification accuracy on validation.
Temporal Event Detection in Audio — Dual-path TCN–Transformer architecture for beat tracking, with parallel branches capturing local temporal structure and global sequence dependencies. Compact model matching state-of-the-art with fewer parameters.
Structured Sequence Modelling — Downbeat tracking in Jangdan rhythmic structures using TCN + RoFormer with a domain-aware Dynamic Bayesian Network. Currently developing a generative audio system.