可见-近红外光谱与联合优化策略的孵前种鸭蛋受精信息无损检测  

Non-Destructive Detection of Pre-Incubation Breeding Duck Egg Fertilization Information Based on Visible/Near Infrared Spectroscopy and Joint Optimization Strategy

作　　者：陈灼廷 王巧华[1,2] 王东桥 陈燕斌 李世军 CHEN Zhuo-ting;WANG Qiao-hua;WANG Dong-qiao;CHEN Yan-bin;LI Shi-jun(College of Engineering,Huazhong Agricultural University,Wuhan 430070,China;Key Laboratory of Agricultural Equipmentin Mid-Lower Yangtze River,Wuhan 430070,China)

机构地区：[1]华中农业大学工学院,湖北武汉430070 [2]农业农村部长江中下游农业装备重点实验室,湖北武汉430070

出　　处：《光谱学与光谱分析》2025年第5期1469-1475,共7页Spectroscopy and Spectral Analysis

基　　金：湖北洪山实验室重大项目课题(2022hszd006);国家自然科学基金项目(32372426,32072302);国家重点研发计划重点专项(2022YFD1300400)资助。

摘　　要：种鸭蛋的孵化是鸭蛋和鸭肉生产的重要保障,无精蛋不能孵化出雏鸭,且在孵化箱内容易变质影响受精蛋的孵化。为了解决人工照蛋剔除无精蛋的劳动强度大、资源浪费等问题,以入孵前种鸭蛋为研究对象,提出了一种基于可见-近红外光谱与深度学习的种鸭蛋孵前受精信息无损检测方法。使用可见-近红外光纤光谱仪对321枚樱桃谷种鸭蛋(受精蛋144枚,无精蛋177枚)采集光谱数据,将光谱数据按3∶1的比例划分出训练集和测试集,采用在原光谱数据中添加噪声与随机偏移、随机选取并计算平均光谱两种方法将训练集进行扩充。设计了一个端到端深度学习模型:自动编码1维卷积神经网络CAE-1DCNN,使用卷积、池化层代替自动编码器中的全连接层,得到改进的卷积自动编码器CAE,采用联合优化策略训练CAE-1DCNN模型,使其具备自动编码器在数据的压缩-重构过程中提取有用特征的能力,并且能够有针对性地提取适用于分类任务的特征。采用了竞争性自适应重加权采样算法(CARS)、连续投影算法(SPA)、无信息变量消除算法(UVE)三种常用特征波长选取算法和K-最近邻(KNN)、朴素贝叶斯(NB)、随机森林(RF)三种机器学习分类模型进行组合,与本文提出的模型进行对比;采用t分布随机邻域嵌入算法(t-SNE)将特征提取效果进行可视化。最后采用梯度加权类激活图(Grad-CAM)将本文提出的模型对光谱数据的关注区域进行了可视化,探讨了光谱信息的生物可解释性。研究结果表明,所提出的CAE-1DCNN模型能较好地提取光谱数据中的有效信息,判别准确率为95.06%,可见-近红外光谱技术与深度学习相结合可以实现种鸭蛋孵前受精信息无损检测,使用联合优化策略训练的卷积自动编码器有较好的特征提取能力。端到端的CAE-1DCNN模型便于集成,为开发无损检测设备提供技术支持。The hatching of duck eggs is an important guarantee for producing duck eggs and duck meat.Eggs without sperm cannot hatch ducklings,and they are prone to spoilage in the incubator,which affects the hatching of fertilized eggs.To solve the problems of high labor intensity and resource waste caused by manually removing-sperm-free eggs through egg photography,this paper takes pre-hatching duck eggs as the research object.It proposes a non-destructive detection method for pre-hatching fertilization information of duck eggs based on visible near-infrared spectroscopy and deep learning.This article uses a visible near-infrared fiber optic spectrometer to collect spectral data from 321 Cherry Valley duck eggs(144 fertilized eggs and 177 azoospermia eggs).The spectral data is divided into training and testing sets in a 3∶1 ratio,and the training set is expanded by adding noise and random offset to the original spectral data,randomly selecting and calculating the average spectrum.This article designs an end-to-end deep learning model:the Autoencoder 1DCNN,which uses convolutional and pooling layers instead of the fully connected layers in the autoencoder to obtain an improved convolutional autoencoder CAE.The CAE-1DCNN model is trained using a joint optimization strategy to enable the autoencoder to extract useful features during the data compression reconstruction process and selectively extract features suitable for classification tasks.This article uses three commonly used feature wavelength selection algorithms,namely Competitive Adaptive Reweighted Sampling(CARS),Continuous Projection(SPA),and Uninformative Variable Elimination(UVE),as well as three machine learning classification models,K-Nearest Neighbor(KNN),Naive Bayes(NB),and Random Forest(RF),to combine and compare with the proposed model.The t-distribution Random Neighborhood Embedding(t-SNE)algorithm is used to visualize the feature extraction effect.Finally,this article used a weighted Class Activation Graph(Grad CAM)to visualize the focus areas of spectra

关 键 词：入孵前种鸭蛋 受精信息 深度学习 联合优化 可见-近红外光谱 无损检测 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]