基于机器视觉的蚕豆荚高精度检测方法研究  

Research on high precision detection method of broad bean pods based on machine vision

作　　者：夏子林 张新洲[1] 王文波 夏先飞 陈兰[1] 顾寄南[1] Xia Zilin;Zhang Xinzhou;Wang Wenbo;Xia Xianfei;Chen Lan;Gu Jinan(College of Mechanical Engineering,Jiangsu University,Zhenjiang,212013,China;Nanjing Institute of Agricultural Mechanization,Ministry of Agriculture and Rural Affairs,Nanjing,210014,China)

机构地区：[1]江苏大学机械工程学院,江苏镇江212013 [2]农业农村部南京农业机械化研究所,南京市210014

出　　处：《中国农机化学报》2025年第1期157-163,F0003,共8页Journal of Chinese Agricultural Mechanization

基　　金：江苏省农业科技自主创新资金项目(CX(21)3145)。

摘　　要：蚕豆荚检测是蚕豆自动化采摘的前提与基础,因此,实现蚕豆荚的快速、准确识别与定位十分关键。基于此,提出一种基于深度学习的蚕豆荚检测方法YOLO-B。为增大整体网络的感受野,获取目标丰富的上下文信息从而提高检测精度,提出SPPX模块。由于蚕豆荚特征较为相似,为消除冗余特征,提出Ghost PAN特征融合模块。经过对比试验分析,相比于其他YOLOv7、YOLOv5与YOLOv4,YOLO-B参数量(29.092M)和计算量(95.466G)最低,m AP(92.58%)最高,分别提升1.33%、2.75%与3.74%。在蚕豆荚特征差异较小、生长姿态差异较大的场景下,均能实现准确检测。The detection of broad bean pods is the basis of automatic picking of broad bean,it plays an important role to realize rapid and accurate identification and positioning of broad bean pods.In this paper,a deep learning-based accurate detection method of broad bean pods(i.e.,YOLO-B)was proposed.A SPPX module was proposed to improve the detection accuracy,which could enlarge the receptive field of the whole network and obtain the rich context information of the target.Since the characteristics of broad bean pods were similar,a feature fusion module(i.e.,GhostPAN)was proposed to eliminate redundant features.Compared to other YOLOv7,YOLOv5 and YOLOv4 algorithms,the number of parameter YOLO-B(29.092 M)and computational power(95.466 G)were the lowest,while mAP(92.58%)was the highest,increasing by 1.33%,2.75%and 3.74%,respectively.The accurate detection can be achieved in the scenarios with small differences in pod characteristics and large differences in growth posture.

关 键 词：蚕豆荚 目标检测 深度学习 YOLO 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]