沈阳农业大学学报2025，Vol.56Issue(6)：55-67,13.DOI:10.3969/j.issn.1000-1700.2025.06.006
基于目标检测引导与视觉大模型的番茄叶部病害智能评估
Intelligent Assessment of Tomato Leaf Diseases Based on Object Detection Guidance and Visual Large Models
摘要
Abstract
[Objective]Tomato leaf diseases significantly affect fruit quality and yield,effective detection and severity assessment is crucial for precise control.A smart detection and evaluation technical framework is proposed in this study to tackle the challenges of disease identification and severity assessment in complex field environments.[Methods]The detection and evaluation framework is proposed based on a dual-path architecture.The core of the method is an enhanced object detection model built on YOLOv12,improved by integrating dual-core convolution,dynamic upsampling,and parallel collaborative attention modules.This model first localizes leaves and determines the presence of disease spots,thereby directing the sample to one of two grading paths.For leaves with spots,the detection results are used to prompt the Segment Anything Model(SAM)to segment both the leaf and the spots,enabling severity grading based on the calculated spot area ratio.For leaves without obvious spots,a dedicated grading model incorporating a low-resolution attention mechanism performs expert knowledge-based discrimination.A tomato leaf disease dataset was constructed using self-collected and publicly available data for experimental validation.[Results]The improved detection model increased leaf detection precision and mAP by 3.4%and 3.1%,respectively,achieved a disease type classification accuracy of 99.1%,and improved the mAP for spot detection by 10.74%to 90.1%.Guided by detection boxes,SAM segmented leaves and spots with an average relative error below 9%for area extraction.The grading accuracy for six spotted diseases exceeded 84%(approx.90.5%on average),while the accuracy for samples without distinct spots reached 87.5%,outperforming models like ConvNeXt-T and Swin-T.[Conclusion]The method is effective in leaf and lesion detection,the SAM segmentation based on detection prompts is stable and reliable,and the technical framework is suitable for automated tomato leaf disease detection and grading,presenting a viable tool for intelligent agricultural disease diagnosis.关键词
番茄/叶部病害/病害检测/病害分级/语义分割Key words
tomato/leaf disease/disease detection/severity grading/semantic segmentation分类
信息技术与安全科学引用本文复制引用
周云成,仇城骏,李成镛,张羽,李汪洋,王珏..基于目标检测引导与视觉大模型的番茄叶部病害智能评估[J].沈阳农业大学学报,2025,56(6):55-67,13.基金项目
国家重点研发计划项目(2023YFD1501303,2021YFD1500204) （2023YFD1501303,2021YFD1500204）
