沈阳农业大学学报2025，Vol.56Issue(1)：92-107,16.DOI:10.3969/j.issn.1000-1700.2025.01.010
基于深度学习的植物叶病斑精细化分割方法
Fine Segmentation Method for Plant Leaf Disease Spots Based on Deep Learning
摘要
Abstract
[Objective]In order to solve the problem of poor segmentation accuracy of small target disease spots and disease spot edges in plant leaves,and to achieve accurate assessment of the severity of plant leaf diseases,a refined segmentation method for plant leaf disease spots based on deep learning was developed.[Methods]This article takes dataset I consisting of grape leaf black rot disease,grape leaf black measles disease,and strawberry leaf spot disease,and dataset Ⅱ consisting of apple leaf spot drop disease,apple leaf black spot disease,and apple leaf rust disease as examples.Based on Deeplabv3+,an improved deep learning network called MFA Net is proposed,which uses an improved Xception network as the backbone network.Firstly,a multi-scale feature extraction module was proposed in the encoder section and used to improve the backbone network;This module extracts information of different scales through three branches,and then highlights lesion feature information through a dual branch attention mechanism consisting of coordinate attention mechanism and channel attention mechanism.Secondly,a dual residual cavity space pyramid pooling module was proposed in the encoder section,which uses two residual branches to compensate for the information of the lesion area and utilizes self attention mechanism to help the model capture the detailed information of the input image.Finally,by introducing two fusion modules to construct a decoder,it helps alleviate the problem of information loss and maintain feature richness.[Results]In terms of evaluation indicators,the mIoU of the two datasets were 92.07%and 91.91%,respectively.Compared with models such as Unet,Unet(Resnet50),Unet++,HRNetV2,Deeplabv3+(Resnet101),and Deeplabv3+(Xception),on dataset I,mIoU improved by 3.73%,5.44%,3.18%,2.79%,5.93%,and 2.65%,respectively.On dataset Ⅱ,mIoU increased by 3.82%,5.17%,2.92%,2.38%,6.37%,and 2.13%,respectively.[Conclusion]In the field of plant leaf lesion segmentation,the segmentation performance of this method has been significantly improved,and the segmentation effect of small target lesions and lesion edges has been improved.关键词
多尺度特征/注意力机制/残差结构/小目标病斑分割/病斑边缘分割Key words
mlti scale features/attention mechanism/residual structure/small target lesion segmentation/edge segmentation of lesions分类
植物保护学引用本文复制引用
徐虹,李林峰,杨昊,文武,陈敏,周航..基于深度学习的植物叶病斑精细化分割方法[J].沈阳农业大学学报,2025,56(1):92-107,16.基金项目
四川省科技计划项目(2023JDZH0034) （2023JDZH0034）
四川省自然科学基金项目(2022NSFSC0964) （2022NSFSC0964）
