Zilong: A tool to identify empty images in camera-trap data

摘要

The use of camera traps to research and monitor wildlife results in a large number of images. Many of the images are the result of a false trigger, resulting in an empty photo. Manually removing empty images is time-intensive and costly. To increase image processing efficiency, we present a non-machine learning algorithm to identify empty images in camera-trap data, and developed freely available software, Zilong. We applied Zilong to 53,598 camera-trap images from 24 sites and compared the results to a CNN-based (Convolutional Neural Network) R package MLWIC (Machine Learning for Wildlife Image Classification). Zilong correctly identified 87% of animal images and correctly identified 85% of empty images, while MLWIC identified 65% and 69%, respectively. Our results suggest that Zilong performed better than MLWIC on identifying empty images. Zilong performed well for most of sites (22/24), with reduced performance identifying empty images when there was vegetation swinging significantly in front of camera (2/24). By using Zilong, wildlife researchers can reduce time and resources required to review camera-trap images.