Ecology and biology researchers often use motion triggered cameras to study endangered animals in the wild. The automatic cameras can be strategically positioned to help gather data such as population changes across vast habitats. However, the raw images can come out blurry or the frame may not be useful for study if the system was inadvertently triggered by other animals in the area.  

Because of that, researchers often spend more time categorizing and labeling images than doing actual research with them.  

Cheng Guo, a Ph.D. student at Colorado State University, is developing new algorithms and machine learning techniques to help researchers studying African leopards with that tedious task. Guo has developed a process that can rapidly sort and identify these animals by comparing their individual and distinct spot patterns across a variety of images. It’s similar to the facial recognition technology used with modern phones and apps and has the potential to liberate scientists from a labor-intensive part of their work.   

While the approaches Guo is developing could ultimately be useful for research into other animals like zebras with unique markings, she said several common challenges still need to be overcome. One is generally improving the image processing approaches and deep learning algorithms used to help the computer systems better recognize patterns between two images. Improving on those techniques provides a solid foundation that will help the whole system become more accurate over time at identifying similarities.  

Another potential challenge is the need to include low-quality images captured in the field. Over-exposed pictures, for example, can still offer valuable information for the models to consider when there may be just a few total images to make comparisons to overall. Finding a way to still use those images further expands the available training data – once again improving accuracy of the whole system as it categorizes, and groups related items together automatically.  

Guo said companies that use similar facial recognition technology processes do not have similar constraints because of their access to large amounts of data. Apple, she said, has access to many images of many people’s faces in different conditions – and many more come in every day.

“That information can be used to continually train and improve their facial recognition models,” Guo said. “For our project though, we may only have one or maybe two usable images of a specific leopard to then compare to other animals in the set. That limits the accuracy of the system as it sorts through submissions. It is hard to address though with endangered species that live in large habitats and may not encounter these cameras often or even at all. Every single image we have is very precious for our work.”