AG-Vision: A Dual-Module Approach for Tomato Leaf Disease Diagnosis

Asim Khan^1,2*Samee Ullah Khan²Irfan Hussain¹

• ¹Khalifa University, Abu Dhabi, United Arab Emirates
• ²Advanced Research and Innovation Center (ARIC), Khalifa University, Abu Dhabi, United Arab Emirates

Timely diagnosis of tomato leaf diseases is crucial for precision agriculture. Existing convolutional neural network (CNN) methods excel at localized feature extraction but often struggle to capture the global contextual relationships necessary for robust, real-world diagnosis. We propose AG-Vision, a novel dual-module architecture that strategically fuses local and global visual context. The architecture leverages an EfficientNet-B4 backbone (DeepFolia) for high-fidelity local feature learning, which is seamlessly integrated with a Transformer encoder (VisiLeaf) to model complex global dependencies. Evaluated on the controlled PlantVillage dataset and the challenging field-condition PlantDoc dataset, AG-Vision achieves state-of-the-art performance with real-time efficiency. Specifically, it attained 99.97\% accuracy (F1 99.53\%) on PlantVillage and 96.97\% accuracy (F1 94.47\%) on PlantDoc, with an inference speed of approximately 25 ms per-image. Ablation studies confirm the necessity of both the CNN (DeepFolia) and Transformer (VisiLeaf) modules, positional encoding, and the optimized attention heads for peak performance. Grad-CAM visualizations qualitatively validate that the model correctly attends to disease symptoms. These results demonstrate that combining local and global feature extraction yields a highly accurate and efficient solution, enabling state-of-the-art tomato leaf disease classification suitable for edge deployment in precision agriculture.