Efficient deep learning-based tomato leaf disease detection through global and local feature fusion.

Abstract

In the context of intelligent agriculture, tomato cultivation involves complex environments, where leaf occlusion and small disease areas significantly impede the performance of tomato leaf disease detection models. To address these challenges, this study proposes an efficient Tomato Disease Detection Network (E-TomatoDet), which enhances tomato leaf disease detection effectiveness by integrating and amplifying global and local feature perception capabilities. First, CSWinTransformer (CSWinT) is integrated into the backbone of the detection network, substantially improving tomato leaf diseases' global feature-capturing capacity. Second, a Comprehensive Multi-Kernel Module (CMKM) is designed to effectively incorporate large, medium, and small local capturing branches to learn multi-scale local features of tomato leaf diseases. Moreover, the Local Feature Enhance Pyramid (LFEP) neck network is developed based on the CMKM module, which integrates multi-scale features across different detection layers to acquire more comprehensive local features of tomato leaf diseases, thereby significantly improving the detection performance of tomato leaf disease targets at various scales under complex backgrounds. Finally, the proposed model's effectiveness was validated on two datasets. Notably, on the tomato leaf disease dataset, E-TomatoDet improved the mean Average Precision (mAP50) by 4.7% compared to the baseline model, reaching 97.2% and surpassing the advanced real-time detection network YOLOv10s. This research provides an effective solution for efficiently detecting vegetable pests and disease issues.

Plant phenotyping relevance

トマト葉の病害状態を画像から検出する深層学習モデルを開発し、2つのデータセットで性能検証しており、植物フェノタイピング手法が研究の中心である。

Code and data availability

The paper's tomato leaf disease detection experiments use the public CCMT-derived tomato leaf disease dataset, which the authors state is publicly accessible via a Mendeley Data deposit (with the complete dataset available on request from the corresponding author). No author analysis code, trained model checkpoints, or