基于深度学习的群体种鸭蛋受精信息检测方法

Detection Method for Fertilizing Information of Group Duck Eggs Based on Deep Learning

摘要:

针对我国禽蛋孵化行业以人工方式剔除无精蛋自动化程度低的问题，以孵化5d的群体种鸭蛋为研究对象，利用图像采集装置采集群体种鸭蛋图像，在常用单步多框检测器（Single shot multibox detector，SSD）网络的基础上提出一种改进SSD目标检测算法，并采用该方法对孵化早期整盘群体种鸭蛋中的受精蛋与无精蛋进行识别。利用MobileNetV3轻量化网络作为模型的特征提取网络，可快速高效提取图像特征。结果表明：本文建立的模型对孵化早期群体种鸭蛋中受精蛋与无精蛋的平均识别精度为98.09%、召回率为97.32%、漏检率为0，优于改进前网络模型的96.88%、96.17%、1.04%。本文方法可为种鸭蛋孵化产业相关智能机器人或机械手的研发提供技术支撑。

Abstract:

The method of removing unfertilized eggs in China’s poultry egg incubation industry relies on artificial irradiation of eggs, with low degree of automation. The accurate identification of fertilized eggs in group duck eggs during the early incubation period is the key technology to realize the automation and intelligence of the incubation process. A group of duck eggs hatched for five days was taken as the research object, and the images of the group duck eggs were collected using corresponding image acquisition devices. Based on the commonly used single shot multibox detector (SSD) network, an improved SSD target detection algorithm was proposed to accurately identify the fertilized eggs and non-fertilized eggs in the eggs of early hatching period. Using MobileNetV3 lightweight network as a model feature extraction network to quickly and efficiently extract image features. At the same time, the inverse residual block was used instead of the standard convolution layer in the SSD regression detection network to improve the detection network efficiency. The results showed that the average recognition accuracy of the model was 98.09%, the recall rate was 97.32%, and themissed detection rate was zero. It was better than 96.88%, 96.17% and 1.04% of the network model before the improvement. Therefore, this method can provide a new basis for the research and development of intelligent robot or robot hand related to duck egg incubation industry and accelerate the intellectualization of poultry egg incubation industry.