Specific Pathogen Free Chicken Embryo Images Classification Based On Transfer Learning

At present,biological vaccine preparations are mainly produced by the chicken embryos.Before the influenza vaccination and preparation,the chicken embryo fertility should be detected to avoid contamination and preparation failure in the proliferation and culture of the virus strains.Strict screening of chicken embryos is also required to ensure the successful incubation process in the breeding industry.Therefore,the effective classification of chicken embryo is of significance for vaccine manufacturing and breeding industry.This paper takes 6 kinds of SPF(Specific Pathogen Free)chicken embryo images from 9-11 hatching days as the research object,proposes the classification method of chicken embryo images based on Alex Net fine-tuning transfer.The main work of this paper is summarized as follows.(1)Due to the over-segmentation caused by complex background of chicken embryo image,and over-fitting problem of network training by insufficient samples,this paper has extracted ROI(Region of Interest)through preprocessing and used data enhancement method to increase the diversity of training samples.Based on the knowledge of artificial candling and the characteristics of SPF chicken embryos of 9-11 hatching days,6 kinds of chicken embryos(live,weak,unfertilized,crack,infected,hemolytic embryo)are screened.To resolve the problem of excessive redundant information and variable number of image categories of chicken embryo,the ROI is extracted by preprocessing combination of Canny edge detecting,white balancing,median filtering,Otsu segmenting and mask template contour extracting;and the ROI images of date set is expanded to obtain a total of 30,000 chicken embryo image data sets by 7 augmentation operations,which can provide standard training and test samples for subsequent model training.(2)For the similarity of vascular features between chicken embryo images,and for the similarity between eggshell texture and crack characteristics of crack embryo,this paper analyzes the composition structure and principle of CNN network,and has redefined of CNN model,include the descriptions of the learning rate,activation function,newly-added local response normalization layer and newly-added Dropout layer in the fully-connection layer.At the programming environment of Tensor Flow GPU framework(NVIDIA Ge Force 9500 GT,32 G memory),the validation analysis and experimentation on two kinds of data sets of different resolutions image dataset and two kinds of classifiers show that the redefined CNN with ROI extraction,224×224 size and SVM classifier has an better classification effect with an average classification accuracy of 92.47%.(3)To obtain higher classification accuracy of chicken embryo images,this paper uses 6 kinds of ROI chicken embryo image of 224×224 as model training and test samples,and analyzes two transfer learning methods based on Alex Net model including feature transfer and fine-tuning transfer.In feature transfer,the pretrained Alex Net model on the Image Net dataset is used as a feature extractor to extract the depth features of the chicken embryo image in 4,096 dimensions and to input them into the SVM classifier for classification.In fine-tuning transfer,the Image Net dataset is used for pretraining,and all the initialization parameters of Alex Net model are obtained.The chicken embryo image dataset is used to train and obtain all the finetuning optimization parameters of the model.The experimental results show that the classification accuracy of the fine-tuned transfer has a better accurate rate.The classification accuracy rate of live embryo classification is 96.2%,the accuracy rate of unfertilized embryo 96.2%,the accuracy rate of crack embryo 95.6%,the accuracy rate of infected embryo 95.2%,the accuracy rate of hemolytic embryo 95.6%,the accuracy rate of weak embryo 94.8%,and the average classification accuracy rate is 95.6%,which can meet the classification requirements of SPF chicken embryo images of 9-11 hatching days.