Team: J. Helgren, J. Pastor, A. Singh
In this project we analyze The Marinexplore and Cornell University Whale Detection Challenge, where participants were tasked with developing an algorithm to correctly classify audio clips containing sounds from the North Atlantic right whale.
The focus of our analysis is on the winning entry (by Nick Kridler and Scott Dobson), whose methodology combines contrast-enhanced spectrograms, template matching, and gradient boosting.
Using Python along with the R interface to h2o, we reproduce the winner’s algorithm, explain its multiple components in IPython Notebook tutorials, test the results, and fine tune the classifier.
The Kaggle training set includes approximately 30,000 labeled audio files. The test set includes approximately 54,000 files. Each file encodes a two second monophonic audio clip in AIFF format with a 2000 Hz sampling rate.
Available upon request.
Among the techniques explained in the tutorials (Ipython notebooks), we can highlight:
- Contrast enhancement and noise filtering, to enhance the signal of the whale call in the spectrogram
- [OpenCV Template Matching](http://docs.opencv.org/3.1.0/d4/dc6/ tutorial_py_template_matching.html)
- [Scikit-Image Docs - Module Skimage Exposure](http://scikit-image.org/ docs/dev/api/skimage.exposure.html)
- Scipy Signal - Docs
- The Marinexplore and Cornell University Whale Detection Challenge
- Whale Detection Challenge Code
- Mark A. McDonald and Sue E. Moore, Calls recorded from north pacific right whales (eubalaena japonica) in the eastern bering sea, Journal of Cetacean Research and Management 4 (2002), no. 3, 261–266
- Sandra L Harris Robert J. Schilling, Digital signal processing using Matlab