基于RT⁃BiSeNet的苹果叶片病害实时分割与分级算法  

Apple leaf disease real⁃time segmentation and grading algorithm based on RT⁃BiSeNet

作　　者：黄样 陈继清[1,2] 黄力湘 佘锴蓉 郝科崴 HUANG Yang;CHEN Jiqing;HUANG Lixiang;SHE Kairong;HAO Kewei(School of Mechanical Engineering,Guangxi University,Nanning 530007,China;Guangxi Key Lab of Manufacturing System and Advanced Manufacturing Technology,Nanning 530007,China)

机构地区：[1]广西大学机械工程学院,广西南宁530007 [2]广西制造系统与先进制造技术重点实验室,广西南宁530007

出　　处：《现代电子技术》2025年第7期35-42,共8页Modern Electronics Technique

基　　金：国家自然科学基金项目(62163005);广西自然科学基金项目(2022GXNSFAA035633)。

摘　　要：苹果叶片病害的及时分割与准确分级对于提高苹果产量和质量至关重要。然而,在复杂的环境下,图像容易受到相似颜色背景和不同光照等因素的影响,给叶片和病斑的准确分割带来挑战,进而影响病害分级的准确性。针对此问题,文中提出一种实时语义分割算法RT⁃BiSeNet,用于苹果叶片病害的分割和分级。首先,分别对BiSeNet的上下文路径和空间路径进行重构,在保证实时分割速度的同时提高分割精度;其次,在解码器中融合浅层的特征映射,提高了叶片边缘和小病斑的分割效果。实验结果表明,RT⁃BiSeNet算法的mIoU和mPA分别为94.60%和97.13%,参数量和复杂度降低了85.95%和72.23%,分割速度达到130.20 f/s,优于其他实时分割方法。该算法能从复杂的背景中实时分割出叶片和病斑,然后根据分级标准对病害进行分级,可为实际生产中苹果病害的精准防控和治疗提供技术支持。Timely segmentation and accurate grading of apple leaf diseases are crucial for improving apple yield and quality.However,in complex environments,images are often affected by factors such as backgrounds with similar colors and varying lighting conditions,which poses significant challenges to the accurate segmentation of leaves and disease spots.These challenges subsequently impact the precision of disease grading.In view of this,a real⁃time semantic segmentation algorithm,RT⁃BiSeNet,is proposed for the segmentation and grading of apple leaf diseases.The context path and spatial path of BiSeNet are reconstructed to enhance segmentation accuracy while maintaining real⁃time processing speed first,and then shallow feature mapping is integrated into the decoder to improve the segmentation effect of leaf edges and small disease spots.The experimental results demonstrate that the mIoU(mean intersection over union)and mPA(mean pixel accuracy)of the RT⁃BiSeNet algorithm achieve 94.60%and 97.13%,respectively,while reducing the number and complexity of parameters by 85.95%and 72.23%,respectively.The segmentation speed reaches 130.20 f/s,surpassing other real⁃time segmentation methods.The algorithm can effectively separate leaves and disease spots in real time from complex backgrounds and then grade the diseases according to the established criteria.To sum up,it can provide technical support for accurate prevention and treatment of apple diseases in actual production.

关 键 词：苹果叶片 深度学习 语义分割 BiSeNet 复杂环境 病害分级 实时分割 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]