Foundation vision models in agriculture: DINOv2, LoRA and knowledge distillation for disease and weed identification

Borja Espejo-García, Ronja Güldenring, Lazaros Nalpantidis, Spyros Fountas

Abstract: Highlights•A novel methodology integrating DINOv2, LoRA and Knowledge Distillation was presented.•A new image classification dataset derived from the RumexWeed dataset was presented.•The use of foundation models outperformed any other Self-Supervised alternative.•The use of SSL-pre-trained models outperformed classical IN1k supervised training.•Visualizations of the features extracted by DINOv2 show agricultural comprehension.

External IDs:dblp:journals/cea/EspejoGarciaGNF25