基于深度学习与颜色特征规则的茶树炭疽病病斑量化分析OA

Anthracnose is one of the major foliar diseases that threatens the growth and development of tea plants and compromises tea quality.At present,no standardized criterion has been established in China for evaluating tea resistance to this disease.Conventional identification and evaluation practices are largely adapted from other crops,and the associated methods rely on visual inspection,resulting in strong subjectivity,low efficiency and limited accuracy.These shortcomings severely constrain both the progress and standardization of elite tea cultivar breeding.In this study,we developed an automated evaluation framework based on deep learning and image analysis.First,Mask R-CNN was used to localize and perform instance segmentation of diseased leaves from field images.Subsequently,within the leaf region of interest(ROI),fine lesion segmentation was achieved by integrating HSV color-space thresholding rules with the Excess Green(ExG)index.Based on the lesion area ratio,a 0-9 severity grading standard was established,and an automated evaluation system was developed.The results show that the model achieved mean of intersection over union(mIoU)of 92.56%for leaf mask segmentation and ratio of average precision at IoU=0.5(RAP@0.5)of 98.91%for leaf instance detection.For consistency validation of lesion detection,the lesion area ratios measured by the automated method were highly significantly correlated with manual reference values generated using the Fiji-Weka tool(Pearson's r=0.946,P<0.001),meeting the accuracy requirement to replace subjective manual grading.