A new computer program can identify bears by their facial features, a leap forward for both facial recognition technology and conservation efforts. The program could have applications on Kodiak for identifying bears in the wild, as well as those that wander into town, according to a local biologist.
Researchers with conservation and software development backgrounds teamed up to create the program, called BearID. The team that designed it includes Melinda Clapham, a conservation biologist at the University of Canada, and software developers Ed Miller and Mary Nguyen.
BearID can do two rather remarkable things, both with a good deal of success.
One, it can pick out a bear’s face in a picture. According to a research paper published on Nov. 6, the program can pick out a bear’s face from a background of a river, trees or other habitat close to 98 out of 100 times. Second, BearID can tell one bear from another 84% of the time.
To do this, the researchers collected 4,675 pictures of 132 different bears. Seven photographers and National Park Service staff at Katmai National Park donated photos from the Brooks River, a noted bear haven. Others came from Clapham and other naturalists who shot pictures of bears at Knight Inlet in British Columbia, Canada. The photos were in differing light and taken between 2009 and 2017.
Nguyen and Miller then modified an existing program designed to find dog faces to pick out bear faces by finding the eyes, nose, ear tips and forehead of the bear. Sometimes, it didn’t match exactly, and the programmers had to go through and manually adjust the face landmarks, the pair wrote on their blog.
This helped the program understand how to pick out and identify different bear features. The team used 3,740 bear faces to “train” BearID.
Then, they set it loose on the remaining 935 photos, and 84% of the time, it was able to recognize different photos of the same bear.
This is not the first time biologists and programmers have teamed up to identify animals in the wild. Other researchers have designed applications that can pick out chimpanzees with 93% accuracy or giant pandas with 96% success. But, as the BearID team noted in their
research paper, these animals have distinct patterns of markings around their faces.
Bears do not have such distinct facial markings, and bears also gain and lose hundreds of pounds of weight in any given year, making them even harder to identify. That the program can work on bears means that it might be able to be used on other species that lack clear identifying features, such as face markings, stripes or spots.
The striking thing about the program is that it’s not just mimicking what Miller and Nguyen did when they manually picked out bear features. It is determining for itself what makes each different from each other, using the manually input data as a starting point. This is a computer programming technique called “deep learning.”
BearID could have many uses in the wild, including on Kodiak Island.
“Identifying individual bears has been an issue for us on Kodiak when dealing with bears in town, so having the ability to identify these individuals would be helpful,” Alaska Fish and Game biologist Nate Svoboda wrote in an email.
“In addition, this type of technology could potentially be used for management and research purposes as well. For example, the ability to identify individual animals at different times could be useful when conducting ‘mark-recapture’ studies to estimate brown bear populations.”
In another blog post, the researchers said the application could be a time saver for anyone working with bear populations.
“The BearID application will automatically scan images and videos from camera traps and citizen scientists to find and identify any bears present in the data. This will vastly reduce the amount of time bear researchers need to spend manually analyzing data,” they wrote.
Going forward, BearID will only get better. It’s an open-source program, so anyone can use it if they have the know-how. As more bear photographs get added to it, the ability of the deep-learning algorithm to pick out bears will improve.