A general-purpose edge-feature guidance module to enhance vision transformers for plant disease identification

Abstract

As agricultural applications are specialized, plant diseases are diverse, and there is a lack of agricultural datasets, current plant disease identification performance is inadequate. In this study, vision transformer (ViT)-like methods are applied to plant disease identification, and an edge-feature guidance module (EFG) to enhance insufficient local information, such as edge features, is proposed for the first time. The consistency and performance of the ViT-based EFG module is verified on four datasets. In particular, the EFG module is relatively independent and can improve the feature extraction capabilities of any ViT-like method. We design efficient ViT backbones and combine them with state-of-the-art methods, namely, ViT, PVT, and Swin, to enhance the fusion of multiscale features and edge information. Our proposed approach is particularly effective for improving plant disease identification. The results of comparative experiments on Paddy, Wheat, Cabbage, and Coffee datasets demonstrate that the proposed method improved feature-fitting performance and outperforms other state-of-the-art models. Code is available at https://doi.org/10.24433/CO.4873751.v2.