The paper introduces the Mix2 framework to tackle multi-label imbalanced classification issues in bioacoustics, focusing on classifying anuran species sounds using the AnuraSet dataset. The dataset presents challenges like class imbalance and multi-label instances. The study explores mixing regularization methods like Mixup, Manifold Mixup, and MultiMix to improve classification performance, especially for rare classes. Results show that alternating between these methods during training leads to significant improvements. The proposed Mix2 system dynamically selects from these methods at each training iteration, promoting robustness and generalization. The study evaluates the performance using macro F-score and MobileNetV3-Large architecture trained from scratch with different augmentation techniques.
To Another Language
from source content
arxiv.org
Głębsze pytania