Deep Multiple Instance Learning Ensemble for the Acoustic Detection of Tropical Birds
| Autor: | Jorge Castro, Roberto Vargas-Masís, Danny Alfaro Rojas |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
0106 biological sciences
Artificial neural network Computer science business.industry Deep learning 020206 networking & telecommunications Pattern recognition 02 engineering and technology 010603 evolutionary biology 01 natural sciences Statistical classification Test set 0202 electrical engineering electronic engineering information engineering Spectrogram Artificial intelligence Mel-frequency cepstrum business |
| Zdroj: | ICMLA |
| DOI: | 10.1109/icmla51294.2020.00051 |
| Popis: | Deep learning algorithms have produced state of the art results for acoustic bird detection and classification. However, thousands of bird vocalizations have to be manually tagged by experts to train most of these algorithms. We use three strategies to reduce this manual work: simpler labels, fewer labels, and less labeled data. The Multiple Instance Learning (MIL) approach provides a framework to simplify and reduce the number of labels, as each recording (bag) is modeled as a collection of smaller audio segments (instances) and is associated with a single label that indicates if at least one bird was present in the recording. In this work, we propose an ensemble of deep neural networks based on the MIL framework to predict the presence or absence of tropical birds in one-minute recordings. As only a relatively small number of training observations (1600) are used to train the algorithm, we compare the performance of the individual networks using log mel-scaled spectrogram and mel-frequency cepstral coefficients. The proposed ensemble of deep MIL networks achieved a 0.98 AUC and 0.92 F 1 score performance on the test set, using only a log mel-scaled spectrogram as data representation. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|