LiteAttNet: An Attention-Enhanced Neural Architecture for Tomato Leaf Disease Recognition

Abstract

Tomato leaf diseases can significantly reduce crop yield and quality, creating a need for accurate and efficient automated diagnostic systems. This paper presents LiteAttNet, a lightweight attention-enhanced framework for tomato leaf disease classification. The proposed architecture integrates a MobileNetV3Large backbone with a Squeeze-and-Excitation (SE) attention module to improve feature representation while maintaining computational efficiency. Experimental evaluation on ten disease categories demonstrated superior performance compared with VGG16, DenseNet121, EfficientNetB0, and MobileNetV3Large, achieving 99.2% accuracy, 99.0% precision, 99.0% recall, and a 98.9% F1-score. In addition, LiteAttNet requires only 3.712 million parameters and achieves a real-time inference speed of 14.59 FPS. LIME-based explainability analysis further confirmed that the model focuses on disease-relevant regions, supporting its reliability for practical edge-based precision agriculture applications.