基于改进Cascade Mask R-CNN的花生叶片病斑实例分割研究  

Segmentation of peanut leaf spot instances based on improved Cascade Mask R-CNN

作　　者：袁瑛 李崇 李坤炎 张栋华 曾林鹏 杨文强[1] YUAN Ying;LI Chong;LI Kunyan;ZHANG Donghua;ZENG Linpeng;YANG Wenqiang(School of Mechanical and Electrical Engineering,Henan Institute of Science and Technology,Xinxiang 453003,China)

机构地区：[1]河南科技学院机电学院,河南新乡453003

出　　处：《河南科技学院学报（自然科学版）》2024年第5期29-39,45,共12页Journal of Henan Institute of Science and Technology(Natural Science Edition)

基　　金：河南省科技攻关项目(222102110095)。

摘　　要：针对自然环境下由于花生叶片图像背景复杂而导致检测效果不佳的问题,以花生作为研究对象,提出了一种基于改进Cascade fc-Mask R-CNN实例分割模型研究田间自然环境下花生叶片的病斑分割方法.首先,基于花生叶片病斑分割数据集建立Cascade Mask R-CNN模型;其次,对该模型进行改进,将原来的主干网络替换成ResNext101和特征金字塔的结合,对损失函数进行了调整;最后对掩膜分支进行了改进,构建了Cascade fc-Mask R-CNN模型.将标注好的花生叶片病斑分割数据集输入不同的分割网络模型中进行训练和验证.结果表明,改进的Cascade fc-Mask R-CNN模型精度达到了98.9%,边界框回归精度达到了77.5%,分割精度达到了77.9%.与其他分割模型相比,改进的Cascade fc-Mask R-CNN模型在花生叶片病斑分割数据集上的实例分割识别效果最好.Aiming at the problem of complex background of peanut leaf images in natural environment,which leads to poor detection effect,a method based on the improved Cascade fc-Mask R-CNN instance segmentation model is proposed to study the lesion segmentation method of peanut leaves in natural environment in the field,taking peanut as the research object.Firstly,the Cascade Mask R-CNN model is built based on the peanut leaf spot segmentation dataset,secondly,the model is improved by replacing the original backbone network with a combination of ResNext101 and feature pyramid,and the loss function is adjusted,and then the mask branch is improved to construct the Cascade fc-Mask R-CNN model.The labeled peanut leaf spot segmentation dataset was input into different segmentation network models for training and validation,and after a series of experiments,the results showed that the improved Cascade fc-Mask R-CNN model achieved 98.9%accuracy,77.5%bounding box regression accuracy,and 77.9%segmentation accuracy.Compared with other segmentation models,the improved Cascade fc-Mask R-CNN model has the best instance segmentation recognition on the peanut leaf spot segmentation dataset.

关 键 词：深度学习 机器视觉 实例分割 花生病害 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] S435.652[自动化与计算机技术—计算机科学与技术]