Morphology-aware Tomato Leaf Disease Classification Using Differentiable Voronoi Graph Tokenization Framework

Abstract

Tomato plants are valuable agricultural crops that are susceptible to various diseases that require timely diagnosis to enhance crop quality and production rates.However, existing models use rigid patching, which causes structural fragmentation and discards important structural information.Furthermore, existing models inadequately represent irregular lesion regions, which are essential for differentiating between leaf diseases.To overcome these issues in tomato leaf classification, this paper proposes Differentiable Voronoi Graph Tokenization (DV-GT) framework, which maps image regions into an irregular graph representation by replacing rigid patching in the transformer model.The dual-path architecture enables the model to effectively extract both texture and spatial features using a Convolutional Neural Network (CNN) and Gabor filter.Furthermore, Voronoi tokenization is used to align with high-gradient image regions, allowing the model to differentiate between diseases.The DV-GT framework utilizes a combination of a Graph Convolutional Network (GCN) and a Vision Transformer (ViT) encoder to extract localized morphological structures and long-range, leaf-wide visual relationships.This enables the DV-GT framework to achieve a classification accuracy of 98.44% on the PlantVillage dataset and 89.47% on the PlantDoc dataset.These results indicate that the proposed DV-GT framework differentiates between visually similar classes and classifies tomato leaf diseases into multiple categories.