Materials, Methods & Technologies, Volume 17, 2023

Abstract: Camera traps are used primarily for the monitoring, conservation, and habitat management of wildlife, especially rare and endangered species. They are preferred among ecological community members who are engaged in protecting and studying the animal ecosystem. Camera traps generate a vast number of images over a given period; however, the majority of these images do not contain any animal sightings. These non-informative images, which are referred to as “garbage images” or “empty images,” are irrelevant and should be eliminated during the research process. The manual process of separating these non-relevant images from those containing the desired animal sightings consumes a significant amount of time and resources, which hinders the efficiency of the research process. To address this challenge, we compiled a diverse collection of camera trap images from various locations and created a dataset consisting of 38,669 empty and 35,083 non-empty images. Additionally, we developed dataset filtering software that sorts the images into folders designated for empty and non-empty images. This software enabled the manual selection and classification of images and facilitated the creation of a large, labeled dataset. Evaluating various state-of-the-art deep learning algorithms using our custom dataset revealed that transfer learning with EfficientNet_B2 resulted in greater than 98% accuracy in distinguishing between empty and non-empty images. With these results, we have been developing an automated filtering program that utilizes deep learning models to effectively discard empty images from camera trap datasets.