Deep Learning for Bioacoustic Recognition in Microcontrollers

Acoustic animal monitoring is crucial for biodiversity conservation and management, but conventional methods are extremely labor-intensive. This research presents a Deep Learning-based automatic animal monitoring system for remote wild areas. Due to the lack of power sources and internet connections, the deep models need to run on extremely resource-constrained microcontrollers. This research introduces a generic pipeline for dramatically downsizing the deep models to fit such devices. Furthermore, a new active learning framework is developed to reduce the amount of human input during training time. Beyond standard benchmarks, this research demonstrates the applicability of the proposed methods to real-world conservation biology tasks.