Multi-kernel inception-enhanced vision transformer for plant leaf disease recognition

Abstract

Abstract The timely and precise identification of diseases in plants is essential for efficient disease control and safeguarding of crops. Manual identification of diseases requires expert knowledge in the field, and finding people with domain knowledge is challenging. To overcome the challenge, computer vision-based machine learning techniques have been proposed by the researchers in recent years. Most of these solutions with the standard convolutional neural network (CNN) approaches use uniform background laboratory setup leaf images to identify the diseases. However, only a few works considered real-field images in their work. Therefore, there is a need for a robust CNN architecture that can identify the diseases in plants in both laboratory and real-field conditioned images. In this paper, we have proposed an Inception-Enhanced Vision Transformer (IEViT) architecture to identify diseases in plants. The proposed IEViT architecture extracts local as well as global features, which improves feature learning. The use of multiple filters with different kernel sizes efficiently uses computing resources to extract relevant features without the need for deeper networks. The robustness of the proposed architecture is established by hyper-parameter tuning and comparison with state-of-the-art. In the experiment, we consider five datasets with both laboratory-conditioned and real-field conditioned images. From the experimental results, we see that the proposed model outperforms state-of-the-art deep learning models with fewer parameters. The proposed model achieves an accuracy rate of 99.23% for the apple leaf dataset, 99.70% for the rice dataset, 97.02% for the ibean dataset, 76.51% for the cassava leaf dataset, and 99.41% for the plantvillage dataset.

Plant phenotyping relevance

植物葉の病害状態を画像から認識する新規深層学習手法を提案・比較評価しており、植物フェノタイピング手法が中心である。

Code and data availability

The paper uses five pre-existing public plant leaf image datasets (Apple, Rice, Ibean, Cassava, PlantVillage) that are cited prior work, not paper-specific assets. No author code, trained model checkpoints, or data deposit with a public URL is mentioned anywhere in the supplied blocks; no availability statement exists.

No evidence-backed public reproduction asset is currently recorded.