摘要
针对水稻病虫害检测精度低、速度慢、模型复杂度高、部署困难等问题,改进了YOLOv4目标检测算法,结合轻量化GhostNet网络,提出了一种基于改进的YOLOv4-GhostNet水稻病虫害识别方法:1)利用幻象模块代替普通卷积结构,替换主干特征提取网络CSPDarkNet53,构建GhostNet模块进行图像的特征提取;2)改进YOLOv4网络的加强特征提取部分PANet结构;3)结合迁移学习与YOLOv4网络训练技巧。通过试验将YOLOv4及其MobileNet系列轻量化网络与Faster-RCNN系列网络和SSD系列网络进行对比,结果表明,改进的YOLOv4-GhostNet模型平均准确率达到79.38%,检测速度可达1 s 34.51帧,模型权重大小缩减为42.45 MB,在保持检测精度达到较高水平的同时模型参数量大幅度降低,适用于部署在计算能力不足的嵌入式设备上。
Aiming at the problems of low accuracy,slow speed,high model complexity and difficult deployment in rice pests and diseases detection,the YOLOv4 target detection algorithm was improved.Combined with the lightweight GhostNet network,a rice pests and diseases recognition method based on the improved YOLOv4-GhostNet was proposed.The phantom module was used to replace the ordinary convolution structure,the backbone feature extraction network CSPDarkNet53 was replaced,and the GhostNet module was constructed for image feature extraction.The PANet structure of the enhanced feature extraction part of YOLOv4 network was improved.Transfer learning was combined with YOLOv4 network training skills.YOLOv4 and its MobileNet series lightweight networks were compared with Fast-RCNN series networks and SSD series networks.The results showed that the average accuracy of the improved YOLOv4-GhostNet model was 79.38%,the detection speed was 34.51 frames per second,and the weight of the model was reduced to 42.45 MB.While maintaining high detection accuracy,the number of model parameters is greatly reduced.It is suitable for embedded devices with insufficient computing power.
作者
周维
牛永真
王亚炜
李丹
ZHOU Wei;NIU Yong-zhen;WANG Ya-wei;LI Dan(College of Information and Computer Engineering,Northeast Forestry University,Harbin 150040,China)
出处
《江苏农业学报》
CSCD
北大核心
2022年第3期685-695,共11页
Jiangsu Journal of Agricultural Sciences
基金
国家级大学生创新训练计划项目(41111214)。
