摘要
我国是鸭蛋与鸭肉生产和消费大国,为此每年需要孵化大量雏鸭才能满足生产需要。由于无精蛋在孵化过程中无法孵化出雏鸭,所以尽早地将其剔除可避免资源的浪费。国内鸭蛋孵化行业需在种鸭蛋入孵7天左右使用人工照蛋方式才能将无精蛋剔除。针对人工照蛋效率低且剔除的无精蛋已无食用价值等问题,以入孵前种鸭蛋为研究对象,将可见-近红外透射光谱技术与卷积神经网络相结合,用于入孵前种鸭蛋受精信息的无损鉴别。实验中对采集得到的400~1 000 nm原始光谱信息进行预处理(去除噪声波段和Savitzky-Golay卷积平滑处理)去除噪声与无关信息,使用竞争性自适应重加权算法(CARS)与连续投影算法(SPA)选取特征波长,并将选择的特征波长转换成二维光谱矩阵。二维光谱矩阵既可以表征特征光谱的有效信息,又可以将光谱信息传入到神经网络进行训练。针对光谱数据的特点,网络过深容易造成模型的过拟合,网络过浅则会造成模型的欠拟合。为此构建了一个4层的卷积神经网络(CNN)用于对光谱信息矩阵进行训练,该网络包括3个卷积层和1个全连接层,卷积层用于自动提取光谱二维信息矩阵的有效信息,全连接层通过对卷积层提取的局部特征进行整合进而供输出层决策,此外在卷积神经网络中引入了局部相应归一化、池化和dropout操作可以加速网络的收敛速度并防止模型过拟合。运用SPA提取的特征波长建立的卷积神经网络模型训练集准确率为97.71%,测试集准确率为97.41%,验证集准确率为98.29%;运用CARS提取的特征波长建立的卷积神经网络模型训练集准确率为97.42%,测试集准确率为97.41%,验证集准确率为97.44%,而使用SPA和CARS提取的特征波长建立的传统机器学习模型测试集精度最高仅为87.39%。研究结果表明,利用卷积神经网络与光谱技术相结合可以实现入孵前种鸭蛋受精信息无损鉴别,可为后续开发动态在线检测设备提供高效、无损、快速的技术支持。
China is a big country for the production and consumption of duck eggs and duck meat.For this reason,a large number of ducklings need to be hatched each year to meet production needs.Since the Infertile egg cannot hatch the duckling during the hatching process,it can be avoided by eliminating it as early as possible.At present,it is necessary to carry out artificial photographing eggs in the country for about 7 days after hatching,so that the infertile eggs can be removed.For the low-efficiency and elimination of artificial eggs,there is no edible value problem.In this paper,the duck eggs before hatching were taken as research objects,and the visible/near-infrared transmission spectroscopy combined with the convolutional neural network was used to realize the duck eggs before hatching Non-destructive identification of fertilization information.The method preprocesses the acquired 400~1000 nm raw spectral information(removing the noise band and Savitzky-Golay convolution smoothing)to eliminate noise and extraneous information,and applies the competitive adaptive re-weighting algorithm(CARS)and continuous projection algorithm(SPA)to select the characteristic wavelengths,and converts the selected characteristic wavelengths into spectral two-dimensional information matrix.The spectral two-dimensional information matrix can not only represent the effective information of the characteristic spectrum,but also can transmit the spectral information to the neural network for training.For the characteristics of spectral data,the network is too deep to cause over-fitting of the model,and the shallow network will cause under-fitting of the model.A four-layer convolutional neural network(CNN)is built to train the spectral information,including three convolutional layers and one fully-connected layer.The convolutional layer is used to extract the spectral two-dimensional information matrix automatically.Effective information,the fully connected layer is integrated for the output layer by integrating the local features extracted by the convolutional layer.In addition,the introduction of local corresponding normalization layer,pooling layer and dropout in the convolutional neural network can accelerate the convergence speed of the network and prevent model overfitting.The accuracy of the model training set established by the SPA extracted feature wavelength is 97.71%,the test set accuracy rate is 97.41%,and the verification set accuracy rate is 98.29%.The model training set accuracy rate established by CARS extraction is 97.42%,the test set accuracy rate 97.41%,the verification set accuracy rate is 97.44%,the traditional machine learning model test set established using the characteristic wavelengths extracted by SPA and CARS has a precision of only 87.39%.The results show that the combination of deep learning and spectroscopy can realize the non-destructive identification of the fertilization information of duck eggs before hatching,which can provide efficient,non-destructive and rapid model support for the subsequent development of dynamic online detection equipment.
作者
李庆旭
王巧华
顾伟
高升
马美湖
LI Qing-xu;WANG Qiao-hua;GU Wei;GAO Sheng;MA Mei-hu(College of Engineering,Huazhong Agricultural University,Wuhan 430070,China;Key Laboratory of Agricultural Equipment in Mid-Lower Yangtze River,Ministry of Agriculture and Rural Agriculture,Wuhan430070,China;National Egg Research and Development Center,Wuhan 430070,China)
出处
《光谱学与光谱分析》
SCIE
EI
CAS
CSCD
北大核心
2020年第12期3847-3853,共7页
Spectroscopy and Spectral Analysis
基金
国家自然科学基金项目(31871863)
国家科技支撑计划项目(2015BAD19B05)
公益性行业(农业)科研专项(201303084)资助。
关键词
入孵前种鸭蛋
受精
卷积神经网络
无损检测
可见-近红外光谱
Pre-incubation duck eggs
Fertilization
Convolutional neural network
Nondestructive testing
Visible/near infrared spectroscopy
