Automatic re-identification of terrestrial mammals using deep learning and camera trap images

Terrestrial mammals serve as flagship species for biodiversity conservation. Individual identification in the wild remains a significant challenge due to dense vegetation occlusion, varied animal poses, and the low-quality imagery captured by camera traps. Continuous deployment generates massive volumes of image data, creating an urgent demand for automated, high-precision re-identification methods. However, existing re-identification methods often struggle with complex backgrounds, limited generalization across species, and a limited number of training entities. To address these challenges, we propose the ARNet, a general framework for terrestrial mammal re-identification. By integrating language-vision priors, the framework effectively isolates animal subjects from cluttered environmental noise, such as dense foliage and irregular shadows. Furthermore, ARNet employs a multi-level feature extraction strategy to capture both body structures and skin patterns, ensuring high discriminative power even under partial occlusion or varied poses. Experiments on a multi-species terrestrial mammal dataset demonstrate that ARNet achieves high accuracy and strong generalization. Our framework provides a scalable and non-invasive tool for ecologists, enabling more reliable population estimation and behavioral tracking, both of which are crucial for the long-term conservation management of terrestrial mammals.