基于改进YOLOv5的鹅蛋孵化早期成活性无损检测  

Nondestructive Detection of Early Hatching Activity of Goose Eggs Based on Improved YOLOv5

作　　者：张猛 乔印虎 ZHANG Meng;QIAO Yinhu(College of Mechanical Engineering,Anhui Science and Technology University,Chuzhou,Anhui 233100)

机构地区：[1]安徽科技学院机械工程学院,安徽滁州233100

出　　处：《中国家禽》2023年第11期126-134,共9页China Poultry

基　　金：安徽省教育厅自然科学重大项目(2022AH040238)。

摘　　要：针对人工照检在鹅蛋孵化成活性检测中存在效率低、劳动强度大、误判率高等缺点,试验提出一种基于改进YOLOv5的鹅蛋孵化早期成活性检测模型。设计2种不同采样系统获取孵化4~7d的鹅蛋特征图像,通过旋转、改变对比度等增强技术构建数据集。在YOLOv5模型的基础上,添加SE(Squeeze-and-Excitation)注意力机制模块增强模型特征提取能力,并引入FReLu激活函数替换原有SiLU激活函数,增强空间灵敏度的同时,显著改善图像视觉。使用改进模型对不同采样系统、不同孵化天数的鹅蛋数据集进行训练、测试,并与原始YOLOv5模型对比。结果显示:选取交并比(IoU)为0.5,2种采样系统的改进YOLOv5模型检测平均精度均值(mAP_(0.5))分别为96.5%、86.3%,单个鹅蛋平均检测时长为62.3ms、34.6ms;对孵化4 d、5 d、6 d、7 d的鹅蛋数据集分开测试,改进YOLOv5模型的mAP_(0.5)分别达到了89.6%、95.4%、98.3%、100%,对比原始YOLOv5模型,改进YOLOv5模型的mAP_(0.5)、F1因子分别提高了3.1%、4.1%。研究表明,改进YOLOv5检测模型在保持轻量化的同时,检测精度高、速度快,可实现对鹅蛋孵化早期成活性的无损检测,并为后期基于工业蛋托实现一次性多个种蛋的快速检测提供依据。In this study a modified YOLOv5-based early hatchability activity detection model for goose eggs was proposed to address the drawbacks of low efficiency,high labor intensity,and high misjudgment rate in manual egg inspection.Two different sampling systems were designed to obtain feature images of goose eggs at 4 to 7 days of incubation.Data augmentation techniques such as rotation and contrast enhancement were employed to construct the dataset.The SE(Squeeze-and-Excitation)attention mechanism module was added to the YOLOv5 model to enhance feature extraction capabilities.The original SiLU activation func-tion was replaced with FReLu activation function to improve spatial sensitivity and significantly enhance image visualizations.The improved model was trained and tested on egg datasets with different sampling systems and incubation days compared with the original YOLOv5 model.The results showed that the intersection over union(IoU)threshold was 0.5,the improved YOLOv5 model achieved mean average precision(mAP_(0.5))of 96.5%and 86.3%for the two sampling systems,respectively.The average de-tection time per goose egg was 62.3 ms and 34.6 ms.When tested separately on datasets of eggs at 4,5,6,and 7 days of incuba-tion,mAP_(0.5) of the improved YOLOv5 model achieved 89.6%,95.4%,98.3%,and 100%,respectively.Compared to the original YOLOv5 model,the improved YOLOv5 model showed an increase in mAP_(0.5) and F1 score by 3.1%and 4.1%,respectively.The re-sults demonstrated that the modified YOLOv5 model maintained lightweight characteristics while achieving high detection accura-cy and speed,achieved non-destructive detection of early-stage hatchability in goose eggs,and provided a basis for fast detection of multiple-seed eggs using industrial egg trays in the future.

关 键 词：孵化成活性 鹅蛋 YOLOv5 目标检测 注意力机制 

分 类 号：S817.6[农业科学—畜牧学] TP391.4[农业科学—畜牧兽医]