基于深度学习的烟草图像分割方法研究进展OA

Tobacco image segmentation is a key enabling technology for intelligent pest and disease diagnosis,automatic harvesting,and precision grading.This paper systematically reviewed the latest advances in deep learning-based tobacco image segmentation,summarizing the research from three application perspectives:tobacco leaf disease segmentation,tobacco stem detection and segmentation,and plant-scale segmentation.A focused analysis was conducted on key techniques such as attention mechanisms,loss functions,lightweighting,and knowledge distillation.Existing studies showed that improved U-Net achieves an IoU of 84.93%in disease segmentation,while improved YOLOv8 attained a mAP of 95.40%in tobacco stem detection.However,the field still faced challenges including scarcity of annotated data,severe class imbalance,insufficient cross-domain generalization,and constraints on edge deployment.Future efforts should be directed toward constructing standardized datasets,developing weakly supervised and generative methods,exploring foundation model transfer,and building end-to-end industrial systems.