Toward improving precision and complexity of transformer-based cost-sensitive learning models for plant disease detection

Abstract

Early and accurate detection of plant diseases is crucial for making informed decisions to increase the yield and quality of crops through the decision of appropriate treatments. This study introduces an automated system for early disease detection in plants that enhanced a lightweight model based on the robust machine learning algorithm. In particular, we introduced a transformer module, a fusion of the SPP and C3TR modules, to synthesize features in various sizes and handle uneven input image sizes. The proposed model combined with transformer-based long-term dependency modeling and convolution-based visual feature extraction to improve object detection performance. To optimize a model to a lightweight version, we integrated the proposed transformer model with the Ghost module. Such an integration acted as regular convolutional layers that subsequently substituted for the original layers to cut computational costs. Furthermore, we adopted the SIoU loss function, a modified version of CIoU, applied to the YOLOv8s model, demonstrating a substantial improvement in accuracy. We implemented quantization to the YOLOv8 model using ONNX Runtime to enhance to facilitate real-time disease detection on strawberries. Through an experiment with our dataset, the proposed model demonstrated mAP@.5 characteristics of 80.30%, marking an 8% improvement compared to the original YOLOv8 model. In addition, the parameters and complexity were reduced to approximately one-third of the initial model. These findings demonstrate notable improvements in accuracy and complexity reduction, making it suitable for detecting strawberry diseases in diverse conditions.

Plant phenotyping relevance

イチゴの病害を画像から検出する軽量Transformer/YOLOベース手法を開発・評価しており、植物の病害状態の画像計測が中心である。

Code and data availability

The paper describes a custom strawberry disease dataset (1,000 curated images plus the Afzaal et al. 2021 dataset) and YOLOv8-based models, but no public repository, dataset deposit, or author code URL is provided. The data availability statement only directs inquiries to the corresponding author, and no allowed URLs (

No evidence-backed public reproduction asset is currently recorded.