Research On Egg Embryo Classification Method Based On Deep Learning

Vaccination of influenza vaccine is one of the main approaches for immunization of infectious diseases.At present,the preparation of influenza vaccine is mainly carried out by the viral cultivation in egg embryos.In the process of strain embryo cultivation,the unseparated dead embryos will cause failure of the cultivation.Therefore,the fertility detection and classification of egg embryos is important for the manufacture of influenza vaccine.The artificial egg irradiation method with poor efficiency are still used in the present fertility detection of egg embryos.Thus,automatic fertility detection of egg embryos from images by machine vision has become a focus in research.To realize the fertility detection and classification of egg embryos,a classification method based on deep learning is proposed in this paper.Two different network structures are designed for 5-day embryos and 9-day embryos.A convolutional neural network containing two branches(TB-CNN)is designed for the small dataset of 5-day embryos whose features are difficult to extract.In the first branch,the dataset is pre-trained with the model trained by AlexNet network on the large-scale ImageNet dataset.In the second branch,the dataset is trained on a multi-layer network.The feature extraction capability of the network is enhanced by TB-CNN,which enriches the extracted features and improves the accuracy rate.Another convolution neural network combining channel weighting and joint supervision(SJ-CNN)is proposed for 9-day embryos with instable features and large intra-class differences.The supervision training method jointing Softmax Loss and Center Loss can convergence intra-class features while guaranteeing the intra-class distance,which makes the features more suitable for classification.A channel weighting module is added to the end of each convolution layer to distinguish the importance of different features by learning,which can enhance useful features and weaken the useless features so as to increase the performance of the model.Experiment results show that our method solves the multi-classification problem of egg embryos.Different models for 5-day embryos and 9-day embryos show good generalization ability and high accuracy rate in verification sets.