Applying South Africa's draft R3337 warning criteria to food-package photographs requires evidence on nutrition, ingredients, and product form. A new study posted on arXiv proposes a region-guided evidence extraction method to pull this information from package images, combined with an abstention-aware screening approach that decides when to defer to a selective vision-language model review.
The system is designed to handle the complexity of real-world food packaging, where relevant details may appear in different locations and formats. By guiding extraction to specific regions and allowing the model to abstain when uncertain, the approach aims to improve reliability in automated compliance screening.
The study is described in the arXiv preprint, though full details of the method and evaluation are not included in the abstract. As the only source, it offers an early look at a potential tool for regulatory monitoring of food labels.