Published in Applied Sciences, MDPI, 2026
Downbeat tracking in Jangdan rhythmic structures using a TCN + RoFormer hybrid and a domain-aware Dynamic Bayesian Network. Incorporates prior knowledge of Jangdan cycle lengths directly into the DBN transition model, enabling accurate downbeat inference with limited annotated data.
Published in Applied Sciences, MDPI, 2024
Parallel TCN–Transformer architecture for temporal event detection in audio. TCN handles local temporal patterns with low complexity; Transformer captures long-range sequence dependencies. Combined model matches state-of-the-art beat tracking accuracy with fewer trainable parameters and no data augmentation.
Published in Electronics, MDPI, 2024
This study introduces a method for Pine Wilt Disease detection using drone orthomapped images. YOLOv8 segments diseased areas, followed by cropping regions with context for classification via Deep Metric Learning. A ResNet50 model with semi-hard triplet loss generates embeddings, and a Random Forest classifier identifies tree species and reduces false positives, outperforming multi-class segmentation.