棉田复杂背景下棉花黄萎病病斑分割算法研究  

Study on Segmentation Algorithm of Cotton Verticillium Wilt Disease Spot in Cotton Field Under Complex Background

作　　者：闫靖昆 黄毓贤 秦伟森 高攀[1] Yan Jingkun;Huang Yuxian;Qin Weisen;Gao Pan(College of Information Science and Technology,Shihezi University,Shihezi 832003,China)

机构地区：[1]石河子大学信息科学与技术学院,新疆石河子832003

出　　处：《南京师大学报（自然科学版）》2021年第4期127-134,共8页Journal of Nanjing Normal University(Natural Science Edition)

基　　金：国家自然科学基金项目(61965014);兵团“强青”科技创新骨干人才计划项目(2021CB030).

摘　　要：在利用机器视觉技术进行大田环境下棉花黄萎病的病叶分割与病斑提取过程中,由于光线明暗、与棉花叶片像素相近的杂草的影响,会出现过分割、误分割等情况.针对此问题,本文提出了基于数据迁移的DeepLabv3+模型与质心选择K均值聚类机制相结合的两阶段分割算法.首先,利用基于数据增强的DeepLabv3+分割模型在复杂背景中提取到病叶;然后在叶片HSV颜色空间中选取初始质心,利用K均值聚类算法得到病斑簇;最后利用数据迁移的方法,把从源领域(Kaggle NPDD数据集)学习到的知识迁移到目标领域(棉花病叶),有效缓解了因为样本数据集数据量较少带来的过分割、误分割问题.试验结果表明,棉花病叶的分割综合评价指标值为98.87%,黄萎病病斑的分割综合评价指标值为87.29%.本文提出的病斑分割算法能够有效分割复杂背景图像中出现的棉花病叶、病斑,时效性更强、准确度更高,可为后续棉花病害叶部图像的进一步识别处理提供技术支撑,为农作物病虫害识别技术的发展提供了算法参考.In the process of cotton verticillium wilt segmentation and spot extraction using machine vision technology in the field environment,there will be over-segmentation and missegmentation due to the influence of the light and shade of weeds with similar pixels of cotton leaves.In order to solve the problems,a two-stage segmentation algorithm combining the DeepLabv3+model based on transfer learning and the K-means clustering mechanism of centroid selection was proposed in this paper.First,the infected leaves were extracted from the complex background by the improved DeepLabv3+segmentation model based on deep enhancement.Then,the initial centroid was selected from the HSV color space of the leaves,and the disinfected patches were obtained by the K-means clustering algorithm.Last,the deep transfer learning technology can be used to transfer the knowledge learned from the source domain(Kaggle NPDD data set)to the target domain(cotton infected leaves),and the problems of over-segmentation and missegmentation were effectively alleviated.The results showed that the segmentation index value of cotton infected leaves was 98.87%,and that of verticillium wilt spot was 87.29%.The algorithm proposed in this paper can effectively segment infected cotton leaves and disinfected cotton patches in complex background images,with stronger timing and higher accuracy.It can provide technical support for further recognition and processing of cotton disinfected leaves images in the future,and provide an algorithm reference for the development of crop pest identification technology.

关 键 词：机器视觉 深度学习 语义分割 棉花病害 大田环境 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]