Copy of the body of the article
Have you ever looked at a stamp and thought: Is it the same stamp I saw yesterday? Well, soon there could be an app based on the new facial recognition technology. Known as SealNet, this seal face research system was developed by a team of undergraduate students at Colgate University in New York.
Inspired by other technologies adapted to recognize primates and bears, Krista Ingram, a biologist at Colgate University, led students to develop software that uses deep learning and a convolutional neural network to distinguish a sealing face from another. SealNet is designed to identify the port seal, a species with a predisposition to land on shores in transport.
The team had to train their software to identify seal faces. “I give him a picture, find his face, [and] cut it to a standard size, “says Ingram. But then she and her students would manually identify the nose, mouth and center of the eyes.
For the project, team members made more than 2,000 images of seals around Casco Bay, Maine, over a two-year period. They tested the software using 406 different stamps and found that SealNet could correctly identify the faces of the stamps 85 percent of the time. Since then, the team has expanded its database to include about 1,500 stamp faces. As the number of stamps registered in the database increases, so should the accuracy of identification, Ingram says.
SealNet developers trained a neural network to distinguish port seals using photos from 406 different seals. Photo courtesy of Birenbaum et al.
As with all technology, however, SealNet is not infallible. The software saw seal faces on other parts of the body, vegetation and even rocks. In one case, Ingram and his students made a double take on the strange resemblance between a rock and a seal face. “[The rock] it looked like a seal face, “says Ingram.” The darker parts were about the same distance as the eyes … so you can understand why the software found a face. “Consequently, he says it’s always best to check manually that the seal faces identified by the software belong to an actual seal.
Like a tired seal climbing on a beach for an involuntary photo shoot, the question arises as to why all this is necessary. Ingram believes SealNet could be a useful, non-invasive tool for researchers.
Of the world’s pinnipeds, a group that includes seals, walruses, and sea lions, harbor seals are considered the most scattered. However, knowledge gaps exist. Other techniques for tracking stamps, such as labeling and aerial tracking, have their limitations and can be very invasive or costly.
Ingram points to site fidelity as an aspect of the seal’s behavior that SealNet could shed more light on. Team tests indicated that some harbor seals return to the same transport sites year after year. Other seals, however, such as two animals that the team nicknamed Clove and Petal, appeared together in two different places. Increasing scientists ’understanding of how seals move could bolster arguments to protect specific areas, says Anders Galatius, an ecologist at Aarhus University in Denmark who was not involved in the project.
Galatius, who is responsible for controlling Denmark’s seal populations, says the software “looks very promising.” If identification rates are improved, it could be combined with another method of photographic identification that identifies stamps using distinctive markings on their fur, he says.
In the future, after more testing, Ingram hopes to develop a SealNet-based application. The app, he says, could allow citizen scientists to help record seal faces. The program could also be adapted for other pinnipeds and possibly even for cetaceans.