Taejeon Christian International School
Jeewon Kim
Abstract
Animal welfare evaluation in industrial and labo- ratory settings remains challenging due to reliance on manual inspection, subjective scoring, and inconsistent criteria across facilities. We propose an automated image-based framework that classifies multiple welfare attributes—such as overcrowding, poor hygiene, lack of enrichment, and visible injury—and syn- thesizes these into a continuous severity index. Leveraging the Kaggle “Animal Welfare” dataset, which includes images and associated metadata from variable environments, our method employs imbalance-aware multi-label learning, probability cal- ibration, and visual attribution to enhance both accuracy and transparency. Designed to generalize across housing types and capture severity continuously, this pipeline provides objective, reproducible metrics useful for compliance monitoring, welfare benchmarking, and longitudinal studies.
