Identifying Endangered Whales with Deep Neural Networks

With less than 500 North Atlantic right whales left in the world’s oceans, knowing the health and status of each whale is integral to the efforts of researchers working to protect the species from extinction.

The current process is quite time-consuming and laborious. It starts with photographing right whales during aerial surveys, selecting and importing the photos into a catalog, and finally comparing the photos against known whales in the catalog by trained researchers.

As part of an ongoing preservation effort, NOAA Fisheries launched a Kaggle data science competition to create the best automated process for identifying individual right whales.

Second place finisher Felix Lau describes how he used cuDNN, GeForce GPUs for initial development and an Amazon Web Services GPU instance to train his deep convolutional neural network.

In his blog, Felix highlighted a variety of different approaches he took for the challenge. He mentioned, “the main performance bottleneck (in the other approaches) is that the classifier was not able to focus on the actual discriminating part of the whales (i.e. the callosity pattern). So in this approach, a new aligner replaced the localizer. Particularly aligner rotate the images so the whale bonnet is always right to the blowhead.”