Research On Classification Methods Of Hatching Eggs Based On Multimodal

Influenza virus has the characteristics of high mortality,high morbidity and contagion.The outbreak of influenza seriously endangers human health,and countless people have lost their lives due to the flu.At present,vaccination is a very effective method to prevent the occurrence and spread of influenza.In the preparation of avian influenza vaccine,the current common method is hatching eggs culture.In the hatching eggs culture method,the virus is injected into the 9-days hatching eggs,and then propagated in the chicken embryo allantoic cavity,and cultured until 16 days for separation and purification.16-days hatching egg is the last stage of the cultivation.The necrotic eggs threaten the quality of the vaccine and must be completely removed.At present,industrial production mainly relies on manual screening,which requires a large number of labors.The production efficiency is very low and the phenomenon of wrong selection and missing selection will occur after the workers are tired.Therefore,finding an automated,efficient and intelligent 16-days hatching eggs classification method has become the key task in vaccine production.In order to solve the problem of image feature similarity between different categories of 16-days hatching eggs,this paper proposes a multimodal hatching eggs classification method.Firstly,a system was designed to collect the image signals and heartbeat signals of the hatching egg.And a Butterworth high-pass filter is designed to remove various noises such as power frequency interference and mechanical motion noise of heartbeat signals.Extracting the region of interest(ROI)of the image eliminates the influence of image background and neighboring embryos on category judgment.Secondly,a 16-days hatching eggs classification network is proposed.The Pic Net branch network based on the Convolutional Neural Network(CNN)theory is used for embryo image feature extraction,and the Pic Net branch is pre-trained with transfer learning strategies.Heart Net branch network based on Temporal Convolutional Network(TCN)is used for heartbeat signal feature extraction.Decision layer is based on Recurrent Neural Network(RNN)theory,which can capture dynamic information in serialized hatching egg characteristics.The two features extracted are fused through their respective decision-making layers,and finally the fused features are input to the fully connected layer to predict the egg embryo category.After training and testing on the 16-days hatching egg dataset,the results show that the accuracy rate of the network in the classification of live embryos,recovered embryos and waste embryos is 98.98%,99.04%,and 99.10%,respectively.The experimental results show that the multimodal hatching eggs classification method proposed in this paper uses two kinds of modal signals at the same time,which overcomes the problem of different categories have similar characteristics and shows good performance.