Jujube Defect Recognition Technology And Its Application Research Based On Deep Learning

Jujube,as a medicinal food with high nutritional value and a kind of health food that can enhance human immunity,is deeply loved by consumers.As a major exporter of jujube processed,strict control of jujube fruit quality is of great significance to ensure the sustainable development of the whole jujube industry chain.Therefore,it is necessary to screen jujube before further processing.Nowadays,the selection of jujube mainly depends on manual works.As the largest jujube producing country in the world,the workload of jujube selection is heavy and the efficiency is low.Meanwhile,the screening personnel have direct contact with jujube,which does not satisfy the food hygiene standards.Therefore,in order to improve the efficiency and automation level of jujube screening,this paper deeply and systematically studies the jujube defect recognition method based on deep learning,The main research contents are as follows:Firstly,this paper constructs three jujube datasets for wizened jujube,yellow-skin jujube,mildew jujube,broken-head jujube and normal jujube,which provides effective data support for defect recognition of jujube based on deep learning.This paper studies the jujube defect recognition model based on lightweight convolutional neural network.The model takes the fire module of squeezenet network as the basic module,optimizes and improves it,constructs the IFM module,and uses the parallel structure of the perception module in googlenet for reference,connecting multiple IFM modules in parallel and integrating them into the attention module,The channel and its importance are weighted to enhance the useful feature information and suppress the redundant information,so as to enhance the learning ability of the model.The proposed lightweight convolutional neural network jujube defect recognition model takes less memory,reduces the model parameters while maintaining a high recognition accuracy,which provides technical support for the deployment of the model in Android mobile terminal.The jujube defect feature fusion network model based on VGG16 is studied,and the model is deployed on the computer terminal.The network model uses 1 × 1convolution and global average pooling,using the attention module to adaptively learn the weight of the side branch,effectively complete the jujube defect recognition.The trained model is tested on the test set,and the results show that the model has high recognition accuracy and good generalization.The paper studies the method of jujube defect recognition based on the ensemble transfer learning model and deploys the model on the web terminal.Firstly,we use the Densenet121 model,Mobilenet model,Inception V3 model and Xception model to extract the features,then migrate to the jujube data set,and train four different classification models respectively.Then,ensemble transfer learning model ETLCNN is constructed by weighted average method.The model is trained with jujube data set3,the results show that the ensemble transfer learning model has stronger feature learning ability and well generalization under the condition of less sample data.Based on the characteristics of the jujube defect recognition models and the different deployment environments,the models are deployed and applied in Android mobile terminal,computer terminal and Web terminal respectively.