高光谱漫反射无损检测鸡蛋受精状态

Non-Destructive Detection of Egg Fertilization Status Based on Hyperspectral Diffuse Reflectance

种鸡蛋孵化期间受精状态的检测需要消耗大量人力、物力,并且孵化期间的种鸡蛋不能保证均为健康蛋,需要能够在孵化早期将无精蛋和死精蛋快速准确挑选出来达到降低生产成本的目的。以白来航鸡蛋为研究对象,采用高光谱分选仪批量采集受精、未受精、死精三类鸡蛋共119枚在382~1026 nm范围内的高光谱数据,其中受精蛋采集孵化3,5,7,9,11,13和15 d的数据,并通过黑白校正方法对原始光谱图做校正处理,得到其漫反射率,经过实验对比以及根据实际生产需要,受精蛋选用孵化3和5 d的光谱数据作为建模数据。同时提出了一种将光谱数据转换为图像数据的方法,在最大化保证光谱原始数据的前提下达到了光谱向量数据可视化的效果,可以有效与深度学习图像识别算法相结合。采用连续投影算法(SPA)、竞争性自适应重加权算法(CARS)对光谱波段进行筛选,建立基于全波段、CARS筛选的特征波长、SPA筛选的特征波长与SVM、RandomForest算法与AlexNet、MobileNet网络的判别模型,其中AlexNet-5dFull Wave Bands准确率最高为93.22%。与通过不同特征波长算法筛选出的数据实验结果对比发现,经过SPA算法筛选后的特征波长的建模效果相比于CARS效果更好,其中SVM-SPA3d模型准确率为91.52%,RandomForest-SPA3d模型的准确率为89.83%,AlexNet-SPA3d模型的准确率为89.83%,表明经过SPA筛选后的特征波长能够保存更多关于种蛋信息差异的有效信息。研究结果表明,利用高光谱分选仪对批量种蛋进行漫反射光谱采集,并将黑白校正后的原始光谱漫反射率数据转换为图像数据,将转换后的图像数据利用深度学习图像识别算法对鸡蛋的受精状态进行准确、无损鉴别是可行的,为后续相关自动化的批量检测提供了技术支持。

During the incubation period of the breeding eggs,a lot of workforce and material resources are consumed,and the eggs during the incubation period cannot be guaranteed to be healthy fertilized eggs.It is necessary to quickly and accurately select the infertile eggs and dead sperm eggs in the early stage of the breeding eggs to reduce production costs.We take Bailaihang eggs as the research object and use a hyperspectral sorter to collect 119 fertilized,unfertilized,and dead eggs in batches with hyperspectral data in the range of 382~1026 nm.The original spectrum is corrected by the black and white correction method to obtain the diffuse reflectance of the egg.After experimental comparison and actual production needs,3d and 5d spectral data are selected as modeling data.We also propose a method to convert spectral data into image data,which achieves the effect of visualizing spectral vector data under the premise of maximizing the guarantee of the original spectral data and can be effectively combined with deep learning image recognition algorithms.We use SPA and CARS to filter the spectral bands and establish a discriminant model based on the full band,the characteristic wavelengths filtered by CARS,the characteristic wavelengths filtered by SPA and SVM,the Random Forest algorithm and AlexNet,MobileNet network.The highest accuracy rate of AlexNet-5d Full Wave Bands is 93.22%.By comparing the experimental results of the data after the screening of different characteristic wavelength algorithms,it is found that the modeling effect of the characteristic wavelengths filtered by the SPA algorithm is better than that of CARS.The accuracy of the SVM-SPA3d model is 91.52%.The accuracy of the RandomForest-SPA3d model is 89.83%.The accuracy of the AlexNet-SPA3d model is 89.83%.The results show that the characteristic wavelengths screened by SPA can save more effective information about the difference inbreeding egg information.The research results in this paper show that the diffuse reflectance spectrum values of batches of hatching eggs are collected by a hyperspectral sorter first,and then the original spectral diffuse reflectance data is converted into image data.Combining image data with deep learning image recognition algorithms is feasible to accurately and non-destructively identify the fertilization state of eggs.This study provides technical support for subsequent related automated batch testing.