A deep active learning system for species identification and counting in camera trap images

Norouzzadeh and colleagues built an end-to-end active-learning system that combines automated detection, classification, and counting with strategic human input, asking experts to label only the most informative images to improve the model efficiently. This human-in-the-loop design reduces the annotation burden while maintaining accuracy and adapting to new datasets where pretrained models underperform. The system addresses a central obstacle to deploying machine learning on novel camera-trap projects: the cost of producing enough labeled training data for unfamiliar species and environments. By optimizing which images humans review, it makes automated processing practical for diverse studies. The paper is an important reference for scalable, transferable machine-learning workflows in camera-trap ecology and for active-learning approaches to wildlife image analysis.