Tuesday, July 01, 2008

Penguin-Recognition Software

Remote cameras and software could be a boon to threatened-species monitoring.

By Lissa Harris

Spot the penguin: Bristol University researchers are using object-recognition software to monitor a colony of rare penguins on Robben Island, in South Africa. A green box indicates that the software “sees” a penguin; a yellow box means that the software has identified a particular animal based on its unique pattern of chest spots. Credit: Tilo Burghardt

Scientists studying a colony of rare penguins on a remote South African island are using sophisticated object-recognition software to identify and track individual animals--an approach that they believe could transform conservation fieldwork.

The software involved--originally developed for recognizing individual human faces--has developed rapidly in recent years. But so far, the so-called Penguin Recognition Project, run by Bristol University, in England, is the first large-scale attempt to use this technology to catalogue and monitor an entire population of animals in the field.

Robben Island is home to roughly 20,000 African penguins, a threatened species that has declined by 90 percent in the past century. Bristol scientists set up several cameras along paths well-traveled by the penguins. Software picks up the penguins' fingerprint-like patterns of black and white feathers, and uses these patterns to identify individual animals. By tracking individual penguins through time, scientists can learn how long they live, how frequently they're reproducing, and what times of year they are most vulnerable.

"There are quite a lot of people working on computer vision to try to identify objects in images," says Peter Barham, a physicist and one of the project's chief researchers. "No one has applied that technology to looking for animals." He adds that the approach can potentially be used to track any animal that has individually distinct visual patterns. Many animals, from humpback whales to giraffes, have individually distinct patterns of coloration.

Conventional animal population-monitoring techniques are typically very costly and difficult, and inflict stress on animals. Most population biology studies are done by capturing animals and tagging them, or by following and photographing individuals so that they can be catalogued by hand in visual databases.

The penguin project does the same thing without these downsides and gets better data too. "We can work out monthly survival rates for penguins. Not annual--monthly," Barham says. "You get a frightening amount of data." Right now, the Bristol scientists are installing a system that will permanently monitor the entire island, following the success of prototypes tested over the past four years.

This week, the project's scientists are presenting their research at the British Royal Society's annual summer science exhibition in London.

The penguin-recognition software uses a learning algorithm that gets better the more data it encounters. Using a large collection of penguin photographs, Bristol University computer scientist Tilo Burghardt "taught" the software to identify a penguin-shaped object by its chest outline and stripe, a black band with a characteristic shape. Individual penguins are recognized by the unique patterns of spots on their chests, each of which is described in the system by its distance from all other spots. The detector is robust enough to correctly identify individual penguins even when one or more of the spots are covered, says Burghardt.