基于图像处理的蛋鸡体重估测方法研究  

Study on Estimation Model of Layer Weight Based on Image Processing

作　　者：郭蓓佳 籍颖[1,2] 锡建中 周荣艳 陈辉[4,5] 王德贺 GUO Beijia;JI Ying;XI Jianzhong;ZHOU Rongyan;CHEN Hui;WANG Dehe(College of Information Science and Technology,Hebei Agricultural University,Baoding,Hebei 071001;Hebei Key Laboratory of Agricultural Big Data,Baoding,Hebei 071001;Graduate School,Hebei Agricultural University,Baoding,Hebei 071001;College of Animal Science and Technology,Hebei Agricultural University,Baoding,Hebei 071001;Key Laboratory of Broiler and Layer Breeding Facilities Engineering,Ministry of Agriculture and Rural Affairs,Baoding,Hebei 071001)

机构地区：[1]河北农业大学信息科学与技术学院,河北保定071001 [2]河北省农业大数据重点实验室,河北保定071001 [3]河北农业大学研究生学院,河北保定071001 [4]河北农业大学动物科技学院,河北保定071001 [5]农业农村部肉蛋鸡养殖设施工程重点实验室,河北保定071001

出　　处：《中国家禽》2021年第11期68-73,共6页China Poultry

基　　金：山东省重大创新工程项目(2019JZZY020611);河北农业大学理工基金项目(ZD201702、LG201706);财政部和农业农村部国家现代农业产业技术体系(CARS-40)。

摘　　要：为了有效监测蛋鸡育成期的体重,并克服传统体重称量方法造成的蛋鸡应激反应致其生产性能下降的问题,试验采用图像处理技术对蛋鸡体重估测方法进行研究。试验采集6~20周龄罗曼灰蛋鸡的俯视图像、实际胫长和体重数据,对蛋鸡俯视图像进行预处理,经计算提取出目标特征,通过一元拟合分析、逐步分析法和MLP神经网络方法将特征参数和实测体重进行拟合,建立多种体重估测模型,并进行估计准确性的比较。结果显示:一元模型中,二次多项式模型的拟合效果最好,决定系数(R^(2))为0.841,平均相对误差为11.42%;通过逐步分析法拟合的模型R^(2)为0.901,平均相对误差9.23%;MLP神经网络模型的R^(2)达0.960,平均相对误差为4.29%。表明使用MLP神经网络模型拟合的结果最为精确,可作为一种有效的估测蛋鸡体重的方法。In order to effectively monitor the body weight of laying hens during the growing period and overcome the problem of stress reaction caused by traditional weighing methods,the image processing technology was used to evaluate the method of estimating the body weight of laying hens.In the experiment,the top view image,actual tibia length and body weight data of Lohmann Grey layers at 6 to 20 week of age were collected,and the top view image of layers was preprocessed.The target features were extracted by calculation.The feature parameters and measured body weight were fitted by one-way fitting analysis,step-by-step analysis and MLP neural network method.A variety of weight estimation models were established,and the estimation accuracy was compared.The results showed that among the univariate models,the quadratic polynomial model had the best fitting effect,and the coefficient of determination(R^(2))was 0.841,the average relative error was 11.42%;R^(2) of the model fitted by stepwise analysis was 0.901,with an average relative error of 9.23%;R^(2) of MLP neural network model was 0.960 and the average relative error was 4.29%.It′s suggested that the MLP neural network model was the most accurate,which could be used as an effective method to estimate the weight of layers.

关 键 词：图像处理 蛋鸡 体重估测 投影面积 MLP神经网络 

分 类 号：S831[农业科学—畜牧学]