AG-Vision: a dual-module approach for tomato leaf disease diagnosis

Accurate and timely identification of tomato leaf diseases is critical for precision agriculture. Although convolutional neural networks (CNNs) perform well in extracting local visual features, they often lack the ability to model global contextual relationships, limiting robustness in real-world field conditions. To overcome this challenge, we propose a hybrid architecture that jointly learns local and global representations. We present AG-Vision, a dual-module framework that integrates an EfficientNet-B4 CNN backbone (DeepFolia) for fine-grained local feature extraction with a Transformer encoder (VisiLeaf) to capture long-range global dependencies through self-attention. The architecture incorporates positional encoding and optimized attention heads to enhance spatial awareness. AG-Vision was evaluated on the controlled PlantVillage dataset and the real-world PlantDoc dataset. Ablation studies assessed the contribution of individual components, and Grad-CAM visualizations were used to analyze model interpretability. AG-Vision achieved state-of-the-art performance on both datasets, obtaining 99.97% accuracy and an F1-score of 99.53% on PlantVillage, and 96.97% accuracy with an F1-score of 94.47% on PlantDoc. Despite its high accuracy, the model maintained real-time efficiency with an average inference time of approximately 25 ms per image. Ablation experiments confirmed the importance of combining CNN and Transformer modules, positional encoding, and optimized attention mechanisms. Grad-CAM results demonstrated that the model consistently focuses on disease-relevant regions. The findings confirm that fusing local and global feature learning significantly enhances classification accuracy and robustness under diverse conditions. AG-Vision offers an efficient and scalable solution suitable for edge deployment in precision agriculture.