基于可见/近红外光谱和深度学习的早期鸭胚雌雄信息无损检测

胚蛋雌雄识别一直是家禽业发展的瓶颈问题，在禽肉生产过程中倾向于养殖雄性个体，而禽蛋生产产业倾向于养殖雌性家禽。若能在孵化过程中较早鉴别出种蛋的雌雄，不仅能够降低家禽孵化产业的成本，还能够提高禽蛋和禽肉生产行业的经济效益。该文以种鸭蛋为研究对象，为了在种鸭蛋孵化早期实现对种蛋的雌雄识别，构建了可见/近红外透射光谱信息采集系统，在200～1 100 nm的波长范围内采集了345枚孵化了0~8 d的种鸭蛋光谱数据。搭建了适用于种鸭蛋光谱信息的6层卷积神经网络（convolutional neural network, CNN），其中包括输入层、3个卷积层、全连接层与输出分类层。卷积层可以提取光谱中的有效信息，全连接层通过对卷积层提取的局部特征进行整合供输出层分类决策。另外在卷积神经网络中引入局部响应归一化和dropout操作能够加快网络的收敛速度。利用该卷积神经网络构建鸭胚雌雄信息识别网络，通过对比与分析不同孵化天数的识别效果，发现孵化7d的识别效果最佳。随后将孵化7 d的种鸭蛋原始光谱数据进行噪声去除，选取500~900 nm波段用于后续的特征波长选取和建模。分别运用了竞争性自适应重加权算法(CARS)、连续投影算法( SPA)与遗传算法(GA)选择能够区分鸭胚性别的波长点，将选取的特征波长转换为二维的光谱信息矩阵，二维光谱信息矩阵保留了一维光谱的有效信息，同时极大地方便了与卷积神经网络的结合。利用二维光谱信息矩阵和卷积神经网络相结合，实现孵化早期阶段鸭胚的雌雄识别。经检验，基于 SPA算法和CNN网络建立的模型效果较佳，其中训练集、开发集及测试集的准确率分别为93.36%，93.12%和93.83%；基于GA算法和CNN网络建立的模型效果次之，训练集、开发集及测试集的准确率分别为90.87%，93.12%和86.42%；基于CARS算法和CNN网络建立的模型的训练集、开发集及测试集的准确率分别为84.65%，83.75%和77.78%。研究结果表明基于可见/近红外光谱技术和卷积神经网络可以实现孵化早期鸭胚胎雌雄的无损鉴别，为后续相关自动化检测装置的研发提供了技术支撑。

Non-Destructive Detection of Male and Female Information of Early Duck Embryos Based on Visible/Near Infrared Spectroscopy and Deep Learning

Gender identification of embryonated eggs in China has always been a key issue in poultry industry development. In poultry meat production, males tend to be bred, while the egg production industry tends to breed females. If the male and female eggs can be identified in the early hatching process, it will reduce the cost of poultry hatching industry improve the economic benefits of poultry egg and meat production industry. This paper takes duck eggs as the research object. To realize the gender identification of duck eggs at the early hatching stage, a visible/near-infrared transmission spectrum acquisition system was constructed, which can collect the Spectral data of 345 duck eggs hatching from 0 to 8 days with the wavelength range of 200~1 100 nm. A 6-layer Convolutional Neural Network (CNN) for duck eggs’ spectral information was established, including input layer, 3 convolutional layers, 1 fully connection layer and output classification layer. The convolutional layer is used for extraction for the effective information in the spectrum. The full connected layer can integrate the local features extracted by the convolution layer for the classification decision of the output layer. In addition, the introduction of local response normalization and dropout operations in the convolutional neural network can accelerate the convergence speed of the neural network. The convolutional neural network was used to construct a duck embryo male and female information recognition network. By comparing and analyzing the recognition effects of different incubation days, it was found that the recognition effect was the best after 7 days of incubation. Subsequently, the duck eggs’ original spectral data hatched for 7 days were removed for noise, and the 500~900 nm band was selected for subsequent characteristic wavelength selection and modeling. Competitive adaptive reweighting algorithm (CARS), successive projections algorithm (SPA) and genetic algorithm (GA) were used to select the characteristic wavelengths that can distinguish the sex of duck embryos, and the selected characteristic wavelengths are converted into a two-dimensional spectral information matrix. The two-dimensional spectral information matrix retains the effective information of the one-dimensional spectrum and greatly facilitates the combination with the convolutional neural network. They were using a two-dimensional spectral information matrix combined with a convolutional neural network to achieve male and female identification of early hatching duck embryos. After testing, the model based on the SPA algorithm and the CNN network has a better effect, among the accuracy of the training set, development set, and test set are 93.36%, 93.12%, and 93.83%, respectively; the model based on the GA algorithm and CNN network was followed. In other words, the accuracy of the training set, development set and test set are 90.87%, 93.12%, and 86.42%, respectively; the accuracy of the training set, development set and a test set of the model based on the CARS algorithm and CNN network is 84.65%, 83.75%, 77.78%. The research results show that the visible/near-infrared spectroscopy technology and convolutional neural network can realize non-destructive identification of male and female duck embryos in early hatching, which provides technical support for developing subsequent related automated detection devices.