摘要
家蚕计数与体长测量是在家蚕养殖过程中的必要环节,传统家蚕计数及体长测量方法主要人工完成,易受主观因素影响,较难实现对家蚕数量和家蚕体长的快速、准确监控。本文使用深度学习的方法实现了家蚕计数及家蚕体长测量,以饲料育家蚕为研究对象,构建了家蚕关键点检测数据集,提出了YOLOv8-Pose-GE算法。该算法在YOLOv8-Pose的Backbone部分加入GAM注意力机制,可以放大全局交互,进行多层感知器的3D排列,提高模型特征提取能力的同时减少信息损失;在Neck部分添加ECA注意力机制,具有实现全局空间信息聚合的部分和进行跨通道交互进行建模的部分,可以提升模型对重要特征的感知能力,使模型更好的处理提取家蚕关键点特征。YOLOv8-Pose-GE的mAP、P和R分别为94.7%、95.31%和87.98%,均优于其他常用的关键点检测算法。该算法同时兼顾了速度,其FPS达到37.61 s^(−1)。本方法可以依靠YOLOv8-Pose-EG的head部分输出的坐标来对家蚕及家蚕关键点位置进行定位,并按顺序依次用直线连接家蚕关键点,由连线长度得到家蚕体长,同时实现家蚕计数。本文对家蚕拍摄录像中随机截取10帧图片进行计数实验,其MAE_L、MRE_L和MSD_L由分别为1.6头、3.6%和2.1头,说明模型具有较高的准确性的同时具有较高的稳定性。本文对40头家蚕(1-5龄家蚕中各随机取8头)进行测量实验,由结果分析得,该算法具有家蚕龄期越高,测量效果越好的特点,尤其5龄,MAE_L、MRE_L、MSD_L和PCC分别为12.29 px、1.87%、4.15px和0.977,总体误差较小。该算法满足家蚕计数及体长测量的需要,为提高家蚕养殖的质量,加强家蚕品种选育提供技术支持。
Silkworm counting and body length measurement are essential processes in silkworm breeding.Traditional,they are mainly performed manually,which is easy to be influenced by subjective factors,and rather challenging to realize the rapid and accurate monitoring of silkworm counting and body length measurement.In this paper,we employ deep learning method to achieve silkworm counting and body length measurement.Taking feed-breeding silkworms as the research object,we construct a silkworm keypoint detection dataset and propose the YOLOv8-Pose-GE algorithm.This algorithm introduces GAM to the backbone of YOLOv8-Pose,amplifying global interactions,performing 3D alignment of multi-layer perceptrons,and enhancing the model's feature extraction capability while minimizing information loss.Additionally,adding ECA mechanism to the Neck part,which facilitates the aggregation of global spatial information and cross-channel interactions for modeling,improves the model's ability to perceive crucial features,thereby enable it to better process and extract silkworm keypoint features.The YOLOv8-Pose-GE achieves 94.7%mAP,95.31%Precision,and 87.98%Recall,outperforming existing keypoint detection methods.Moreover,the algorithm also balances speed,achieving an FPS of 37.61.Utilizing the coordinates output from the head part of YOLOv8-Pose-GE,this method can locate the position of the silkworms and their keypoints,connect the keypoints of the silkworms with straight lines in order,and obtain the body length of the silkworms from the length of the connecting lines,achieving silkworm counting simultaneously.In this paper,10 frames of randomly intercepted images from the video shooting of silkworms are counted for the experiment,and their mean absolute error(MAE_C),mean relative error(MRE_C)and mean squared deviation(MSD_C)are 1.6 individuals,3.6%and 2.1 individuals,respectively,which indicates that the model exhibits both high accuracy and stability.This study conducts measurement experiments on 40 silkworms(randomly selecting 8 from each of the instar stages1-5).The results show that the algorithm has the characteristic that the higher the silkworm instar,the better the measurement effect.Specifically for the 5th instar,the algorithm's MAE_L,MRE_L,MSD_L,and PCC in fifth instar are 12.29 px,1.87%,4.15px and 0.977,respectively,indicating a relatively small error.This method satisfies the needs of silkworm counting and body length measurement and can provide technical support to improve the quality of silkworm breeding and strengthen the selection and breeding of silkworm varieties.
作者
刘莫尘
孙崇凯
李正浩
常昊
尚明瑞
宋占华
刘贤军
孙廷举
闫银发
LIU Mo-chen;SUN Chong-kai;LI Zheng-hao;CHANG Hao;SHANG Ming-rui;SONG Zhan-hua;LIU Xian-jun;SUN Ting-ju;YAN Yin-fa(College of Mechanical and Electronic Engineering/Shandong Agricultural University,Tai'an 271018,China;Shandong Higher Education Institution Future Industry Engineering Research Center of Intelligent Agricultural Robots,Tai'an 271018,China;Shandong Key Laboratory of Intelligent Production Technology and Equipment for Facility Horticulture,Tai'an 271018,China;Juxian HAITONG Cocoon SILK Industrial Company Limited,Ri'zhao 276800,China;Shandong Guangtong Silkworm Eggs Group,Wei'fang 261000,China)
出处
《山东农业大学学报(自然科学版)》
北大核心
2025年第4期616-627,共12页
Journal of Shandong Agricultural University:Natural Science Edition
基金
山东省重点研发计划项目(2022TZXD0042)
国家蚕桑产业技术体系项目(CARS-18)
山东省蚕桑产业技术体系建设项目(SDAT-18-06)。
关键词
家蚕
深度学习
计数
体长
关键点检测
Silkworm
deep learning
count
body length
keypoint detection
