一种改进的基于Mask R-CNN的玉米大斑病实例分割算法  

An improved Mask R-CNN based instance segmentation algorithm for maize Northern Leaf Blight

作　　者：朱宇浩 童孟军[1,2] Zhu Yuhao;Tong Mengjun(College of Mathematics and Computer Science,Zhejiang A&F University,Hangzhou 311300,China;Zhejiang Provincial Key Laboratory of Forestry Intelligent Monitoring and Information Technology Research,Hangzhou 311300,China)

机构地区：[1]浙江农林大学数学与计算机科学学院,浙江杭州311300 [2]浙江省林业智能监测与信息技术研究重点实验室,浙江杭州311300

出　　处：《电子技术应用》2024年第5期71-76,共6页Application of Electronic Technique

摘　　要：玉米作为我国主粮作物,其生产常受大斑病、小斑病、锈病等病害及虫害影响,导致其产量与品质下降,威胁农业生产安全。近年来,视觉检测技术因其高准确性已成为病害防控的重要工具。以Mask R-CNN为基础框架,通过融入DyHead、Groie和OHEM模块进行优化,旨在提升对细微病灶图像的分割效能。改良后的模型在病害图像分割任务上展现出卓越性能,平均精度(mAP)提升4%,尤其在小目标分割上准确率提高8.5%,相较于YOLOv5、YOLACT++等同类模型优势显著。通过消融实验验证了各新增模块的有效性,证实该模型为精准检测玉米大斑病提供了有力的技术支持与理论依据。Maize,a crucial staple crop in China,is frequently beset by production challenges stemming from diseases such as maize Northern Leaf Blight,Southern Corn Leaf Blight,and rust,along with insect pests.These maladies significantly undermine maize yield and quality,presenting a potential menace to agricultural production stability.In recent times,visual disease detection techniques have emerged as pivotal instruments for disease management,offering heightened precision relative to conventional methods.This paper leverages the Mask R-CNN architecture as its foundation,integrating DyHead,Groie,and OHEM modules to augment the model's proficiency in segmenting images containing minute disease manifestations.The enhanced Mask R-CNN model exhibits outstanding performance in disease image segmentation,witnessing a 4%uplift in mean average precision(mAP)and an 8.5%enhancement in accuracy for small object segmentation.Compared to analogous instance segmentation models like YOLOv5 and YOLACT++,this model displays superior prowess.To substantiate the utility of each incorporated module,ablation studies were carried out,revealing their constructive roles.Thus,this methodology furnishes a sturdy theoretical underpinning and technological means for the efficacious and precise detection of maize Northern Leaf Blight.

关 键 词：实例分割 玉米大斑病 Mask R-CNN 计算机视觉 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]