基于YOLOx的马铃薯芽眼检测
YOLOx-based potato bud eye recognition detection
李海庚 1冯全 1杨森1
作者信息
- 1. 甘肃农业大学 机电工程学院,甘肃 兰州 730070
- 折叠
摘要
马铃薯芽眼检测是实现种薯自动切块的关键因素.为快速、准确检测马铃薯种薯芽眼,提高智能切种装置效率,提出一种基于卷积神经网络的马铃薯种薯芽眼检测方法.首先,针对芽眼形态多样的特点,通过图像增广及图像多样化处理建立马铃薯芽眼数据库,以增强检测网络的泛化能力;然后,利用YOLOx对小尺寸目标的高效特征表达能力,构建用于芽眼检测的网络模型,实现马铃薯芽眼的快速准确检测.试验结果表明:YOLOx网络对含有单个和多个无遮挡芽眼的样本,以及含有疤痕、斑点、虫眼和机械损伤的样本均有良好的检测效果.其中,最终检测精准度P为95.86%,召回率R为 96.95%,平均精度均值mAP为 95.37%,检测速度为 42.3 FPS.对比YOLOv3 和YOLOv4 网络模型,YOLOx模型可以同时满足马铃薯芽眼识别检测的精度和速度要求,可为马铃薯智能化切种提供技术支持.
Abstract
Potato bud eye detection is a key factor in realizing automatic seed potato cutting.In order to quickly and accurately detect potato seed bud eyes and improve the cutting efficiency of intelligent seed cutting device,a potato seed bud eye detection method based on convolutional neural network was proposed.Firstly,according to the characteristics of the diverse morphology of bud eyes,a potato bud eye database was established by image augmentation and image diversification processing to enhance the generalization ability of the detection network.Then,using the efficient feature expression ability of YOLOx for small-size targets,a network model for bud eye detection was constructed to achieve rapid and accurate detection of potato bud eye.The results showed that the YOLOx network can achieve good detection results for samples containing single and multiple unobstructed bud eyes,as well as samples containing scars,spots,insect eyes and mechanical damage.Among them,the final detection accuracy P was 95.86%,the recall rate R was 96.95%,the average accuracy mAP was 95.37%,and the detection speed was 42.3 frames/s.Compared with the YOLOv3 and YOLOv4 network models,the YOLOx model can meet the accuracy and speed requirements of potato bud eye recognition detection at the same time.This method can provide technical support for intelligent potato cutting.
关键词
YOLOx/目标检测/卷积神经网络/马铃薯芽眼Key words
YOLOx/object detection/convolutional neural networks/potato bud eyes引用本文复制引用
基金项目
国家自然科学基金地区科学基金(32160421)
国家自然科学基金面上项目(31971792)
甘肃省教育厅产业支撑项目(2021CYZC-57)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据