Animal classification considering infrared few-shot open-set recognition

This paper introduced the novel Infrared Few-shot Open-set Recognition (IFOR) framework for effective nighttime wildlife surveillance using limited infrared images. We focused on the effectiveness of three critical methods: small-scale datasets, domain shift, and open-set recognition for addressing IFOR. Our study reveals that the Transformer-based feature extractor outperforms the CNN-based feature extractor in infrared image classification, achieving 51.04% accuracy in a 5-way, 1-shot setting. ImageNet pre-training consistently yields superior results compared with formula-driven supervised learning. Additionally, meta-learning significantly enhances classification accuracy (by 11.19%) and unseen species identification (AUROC improved by 7.75%) compared to conventional learning methods based on mini-batch learning. These findings contribute to advancing wildlife monitoring techniques in challenging nocturnal environments and offer valuable insights for developing efficient machine-learning models for limited-data scenarios in ecological research.