Archives of Current Research International

Tea leaf diseases can adversely affect leaf quality and production, creating a need for reliable image-based classification methods. Recent studies have explored convolutional neural networks, Vision Transformers (ViTs), and attention mechanisms for tea disease recognition, but performance remains dependent on dataset composition and evaluation conditions. This study evaluates a lightweight attention-enhanced ViT for three-class tea leaf image classification using the TeaLeafNet dataset obtained from Kaggle (Singh, 2026). The dataset contains 600 images, with 200 images each for Green Leaf, Red Rust, and Blister Blight. A pretrained ViT-Tiny model is compared with the same backbone augmented by a Squeeze-and-Excitation (SE) feature-recalibration module. Both models were evaluated under the same preprocessing and optimisation protocol using three random seeds (42, 123, and 2026). The baseline ViT achieved a mean accuracy of 93.33% ± 1.11%, whereas ViT-SE achieved 96.30% ± 0.64%. The mean Macro-F1 increased from 0.9330 ± 0.0112 to 0.9627 ± 0.0065. The SE module added only 18,432 parameters, corresponding to an approximately 0.33% increase over the baseline. Improvement was observed across all three evaluated seeds, indicating a consistent performance gain under the tested conditions. The findings suggest that lightweight feature recalibration can enhance a compact ViT without substantial parameter overhead. However, the limited dataset size and absence of an independent external test set require cautious interpretation and further validation across diverse field conditions.