关键词: BiSeNet, 深度学习, 语义分割, 病害程度, 分级预测, 葡萄（Vitis vinifera L.）, 黑麻疹

Abstract: In order to accurately grade and predict the degree of black measles disease in grapes（Vitis vinifera L.）, a semantic segmentation model was used to separate the leaf and lesion parts. The ratio of lesion area to total leaf area on the same leaf was used as the basis for disease severity grading, and the degree of black measles disease in grapes was predicted. 419 grapes disease images from the PlantVillage public database were accurately annotated and subdivided into three categories： background, leaves, and lesions, and data augmentation techniques were applied to increase sample diversity. Using BiSeNet as the benchmark model and introducing GhostNet as the backbone extraction network for context paths not only maintained a small number of model parameters, but also achieved a significant improvement in accuracy, meeting the needs of disease severity classification prediction. A cumulative atrous spatial pyramid pooling （CASPP） module was proposed to replace the single context embedding module in the BiSeNet model, in order to enhance the multi-scale context information extraction ability of the BiSeNet model and improve the segmentation accuracy of the model. After testing, the average Intersection over to Union of this research model in the test set was 94.11%. When predicting the degree of black measles disease in grapes, the accuracy reached 98.21%, which could accurately predict the degree of black measles disease in grapes.

Key words: BiSeNet, deep learning, semantic segmentation, disease severity, grading prediction, grapes（Vitis vinifera L.）, black measles disease