Hybridizing DINOv2 and EfficientNet-B0: A Novel Approach for Multi-Class Plant Disease Detection and Severity Analysis via Unsupervised Lesion Localization
Year: 2026
Type: conference-paper
Abstract: Accurate and timely diagnosis of plant diseases is critical for ensuring global food security. While deep learning has achieved remarkable success in disease classification, most existing models fail to quantify the infection severity or require labor-intensive pixel-level annotations for segmentation. To address this limitation, we propose a novel Dual-Stream Feature Fusion Framework that synergizes the global semantic capabilities of self-supervised Vision Transformers with the local texture extraction of Convolutional Neural Networks.Our architecture consists of two parallel streams: A partially fine-tuned DINOv2 backbone, where the final four transformer blocks are unfrozen to adapt high-level semantic features to the agricultural domain; and An EfficientNet-B0 stream, employed to capture high-frequency textural patterns such as fungal lesions and discoloration. Feature vectors from both streams are fused via concatenation to form a robust representation for multi-class classification. Furthermore, we introduce an unsupervised lesion localization mechanism that leverages the self-attention maps from the DINOv2 backbone to estimate disease severity without requiring ground-truth segmentation masks.Experimental results on the PlantVillage dataset demonstrate that our hybrid approach achieves superior classification performance, with an accuracy of 99.82% and an F1-score of 99.71%. Additionally, the proposed severity estimation module generates interpretable visual saliency maps, providing farmers with actionable insights into infection intensity. This study presents a scalable, annotation-efficient solution for precise plant pathology analysis.
Institution: COEP Technological University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: hybrid
OpenAlex ID: W7214977415