EfficientNet–ViT Hybrid Framework with Attention-Guided Fusion for Tomato Leaf Disease Classification

Abstract

In this paper we provide a hybrid deep learning framework for accurate classification of tomato leaf diseases. The proposed model integrates EfficientNet and Vision Transformer (ViT), where EfficientNet captures fine-grained texture and color variations, and ViT models long-range spatial dependencies. An attention-guided fusion mechanism is introduced to combine these representations by assigning adaptive importance to each branch. The model is evaluated on an eight-class dataset containing 18,160 images, divided into training (70%), validation (15%), and testing (15%) sets. Data augmentation and transfer learning are applied to improve generalization, while Bayesian optimization is used for hyperparameter selection. The proposed approach achieves 98.6% training accuracy, 96.7% validation accuracy, and 96.2% test accuracy, with an F1-score of 0.962 and an AUC of 0.978. Five-fold cross-validation further confirms stable performance with low variance. Comparative results show that the hybrid model outperforms standalone EfficientNet and ViT architectures. These findings demonstrate that combining convolution-based feature extraction with transformer-based global modeling provides a reliable solution for tomato leaf disease classification in practical agricultural applications.