Open Set Bioacoustic Signal Classification based on Class Anchor Clustering with Closed Set Unknown Bioacoustic...

Area of Science:

• Bioacoustics
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning for bioacoustic classification typically uses closed set recognition (CSR), limiting its effectiveness with real-world data variations.
• Open set recognition (OSR) addresses this by classifying inputs as known or unknown, but its application in bioacoustics, particularly for animal sounds, is underdeveloped.

Purpose of the Study:

• To propose a novel method for open set bioacoustic signal classification specifically for animal sounds.
• To address the limitations of CSR in handling diverse, real-world bioacoustic data.
• To develop a classifier capable of accurately identifying target animal species and recognizing unknown sounds.

Main Methods:

• A novel open set bioacoustic signal classification method based on Class Anchored Clustering (CAC) loss is proposed.
• The method incorporates closed set unknown bioacoustic signals during training by adding an "Unknown" class, utilizing an n+1 cross-entropy loss alongside CAC loss.
• An animal sound dataset comprising 101 species was created for evaluation.

Main Results:

• The proposed method demonstrated superior performance compared to baseline methods across key metrics.
• Performance improvements were observed in the area under the receiver operating curve for detecting both target and unknown classes.
• The method achieved higher classification accuracy for open set signals and target classes.

Conclusions:

• The developed method effectively classifies known animal sounds while accurately recognizing unknown bioacoustic signals.
• This approach advances the application of open set recognition in the field of bioacoustics.
• The findings suggest a significant improvement in the robustness of bioacoustic classification systems for real-world applications.