Using faster region-based convolutional neural network for automatic detection of baleen whale social calls

• DOI: 10.1121/1.5137333
• Publication Date: 2019-11-14T02:17:33Z
• Authors: Jeppe H. Rasmussen; Ana Širović
• Keywords: Spectrogram; Baleen

Using blue and fin whale calls for estimating population density of the two species from passive acoustic data is an active research topic. However, manually analyzing long-term extremely time consuming yet a necessary first step such work. Blue social are highly variable quite similarity, which has resulted in challenges developing automatic detectors these calls. The applicability faster region-based convolutional neural network (Faster R-CNN) method speeding up this detection classification task was explored. A large dataset D 40 Hz southern California used training network: 1378 spectrograms were created 10-s sound clips each containing call. resulting images contrast-enhanced labeled with region interest (ROI) before being applied as images. probability score ROI modified to favor detections within previously measured frequency range Testing shows Faster R-CNN have very low miss- false positive rate both call types thus promising tool detecting classifying baleen