The team’s AI-powered solution identifies sperm whales using computer vision for image processing.

A team called “AI Sensing” won the Capgemini Group’s (EN: CAP) annual Global Data Science Challenge 2020. It developed a solution that analyzes thousands of photos of sperm whales and identifies matches of the individual whales.

The findings help scientists track the whales’ migration routes, protect their natural habitats, and analyze their social structures. (COMPANYNEWSHQ)

Tail flukes are whales’ “fingerprints”

Scientists identify sperm whales by their tell-tale tail flukes – each is a distinctive pattern much like a human’s fingerprints. However, a photo of a sperm whale in a new location has to match with a previous photo of the same whale to track its movement.

This is a very tedious and manual process that has so far consumed hundreds of hours of scientists’ time.

Capgemini decided to replace this legacy method with a solution that harnessed computer vision and machine learning algorithms.

This task, therefore, became the “challenge” for the GDSC 2020.

The solution from Capgemini’s AI Sensing team

The winning system from AI Sensing comprised a pre-trained deep neural network (ResNet 101). The researchers used about 4,500 pictures to train the neural network. These images contained tail flukes from more than 2,200 individual whales.

The model runs on a GPU-based Amazon Elastic Compute Cloud (Amazon EC2) compute cluster. It can process the photos in three hours.

The team used Amazon SageMaker for their system. SageMaker is AWS’s service that helps build, train, and deploy machine learning models quickly in the cloud and at the edge.

The accuracy of the solution is 97.5%.

Scientist is elated

Lisa Steiner, a marine biologist and renowned expert for sperm whale research in the Azores, who endured the legacy system, is enthused with the results.

The Capgemini AI solution correctly matched a sperm whale that scientists could not identify since 2002. The reason: changes in the contours of its tail fluke.

“I found quite a cool match that had not been detected previously, due to some changes in the contour,” said Steiner. “This goes back to 2002, so there were three separate IDs for this animal, instead of one. If I had just clicked match on the first photo, I wouldn’t have found it. I’m loving this program!”

Related Story:     The Norwegian AI Lab Helps Salmon Aquaculture