InViTMixup: plant disease classification using convolutional vision transformer with Mixup augmentation

ABSTRACT

Computer vision-based research is carried out to enhance the crop yield with automation solutions. In Tamil Nādu, tomato cultivation is carried out in many districts, including Coimbatore, and its contribution to the Indian economy, i.e. GDP growth, is significant. This paper proposes InViT Mixup, a novel approach toward image classification of diseased tomato leaves. The effectiveness of convolutional neural network (CNN) in obtaining spatial information and the attention mechanism of vision transformers, along with Mixup data augmentation, are combined, and the model’s generalization in predicting the diseased tomato based on leaf symptoms is evaluated. The proposed convolutional transformer model improvised the classification results of tomato disease in terms of accuracy and training speed compared to pure CNN or transformer-based models. In the proposed work, 10 tomato leaf disease classes from the PlantVillage dataset are used. Top-1 accuracy of 93.5% and Top-5 accuracy of 99.7% are achieved.

CO EDITOR-IN-CHIEF:

• Hsieh, Sun-Yuan

ASSOCIATE EDITOR:

• Lin, Limei

KEYWORDS:

• Transfer learning
• inception V3
• vision transformer
• mixup augmentation

Nomenclature

D	=	Dimension
k	=	Kernel
γ,β	=	Layer normalization
$w_q^T$	=	Learning vector
N	=	Number of elements
q	=	Query
$x_{\mathit{i}}$, $x_j$	=	Raw images
n	=	Sequence length
α	=	Weight

Disclosure statement

No potential conflict of interest was reported by the author(s).