FNE-RTDETR: A lightweight end-to-end model for small-area tomato leaf disease detection and fine-grained classification.

Abstract

Tomatoes, widely cultivated around the world, are not only an important source of daily nutrition but also a crop with economic significance. However, the yield and quality of tomatoes are highly susceptible to environmental and climatic factors. In view of this, timely detection and accurate identification of leaf diseases in tomato plants play a crucial role in maintaining optimal production and quality. Current detection methods rely on the subjective expertise of agricultural specialists, leading to inefficiencies and inconsistent outcomes that fail to meet the requirement for precise and timely disease management. In addition, it remains a persistent challenge to trade off between lightweight modeling and recognition accuracy in natural environments. In this study, FNE-RTDETR model for detecting tomato leaf diseases was introduced, where FasterNet network was used in place of the original backbone network of RT-DETR to integrate lightweight PConv so as to reduce network parameters and computational complexity while retaining detection performance, thereby improving feature extraction efficiency. Additionally, the combination of Deformable attention and AIFI modules enhances the model's fine-grained classification capability for various disease types. Also, the cross-attention mechanism in the decoder was replaced by an efficient multi-channel attention mechanism, which strengthens the model's ability to fuse multi-scale semantic features across spatial dimensions, thereby improving the detection performance of small lesion areas and addressing effectively the challenges of disease detection. Experimental results demonstrate that FNE-RTDETR achieves an mAP50 of 91.5%, outperforming RT-DETR by 4.1%, while reducing parameters by 14.7% and GFLOPs by 9%. Ablation and robustness experiments further validate the model's superior convergence speed and generalization ability. Compared with YOLOv3, YOLOv5, YOLOv6, YOLOv8, YOLOv9, YOLOv10, and various DETR variants, FNE-RTDETR consistently achieves higher accuracy and better lightweight performance across multiple datasets, demonstrating strong potential for practical application in tomato leaf disease detection.

Plant phenotyping relevance

トマト葉の病斑・病害を画像から検出・分類するFNE-RTDETRモデルを開発し、複数モデル・データセットとの比較、アブレーションおよび頑健性実験で技術検証しているため、植物病害状態のフェノタイピング手法が中心である。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.