The Research On Deep Learning For Jujube Defects Recognition Technology

At present,most of the situations of jujube defects detection still depends on manual inspection at home and abroad.Manual method is not only labor intensive,but also inefficient.The quality of jujube sorting is also influenced by workers subjectively and does not meet the food safety requirements in China.Machine vision technology has become a major trend to promote the level of deep processing,and also meets the processing needs of modern agricultural products.Based on deep learning theory,this paper conducts an in-depth and systematic study on the key technology in the detection of jujube defects.The main research contents are as follows:According to the detection requirements of Xinjiang grey jujube,the automatic device for jujube defect detection was designed based on machine vision.The jujube defect detection experimental platform was built.The design of the image acquisition part was completed,including digital image data transmission mode,light source configuration method,industrial camera and lens model were determined.According to the full surface information detection requirements of jujubes,the structure for conveying and turning device was designed.According to the characteristics of jujube defects,the algorithm of jujube defect detection based on Blob analysis is proposed.The method of separation of jujube and its background were proposed by using color space analysis and Blob analysis.The color space model and the segmentation threshold achieved the defects detection of typical jujubes such as skin crack fruit,mildewed fruit and starch head fruit quickly and accurately.In order to solve the dried immature fruit and yellow skin fruit defects inaccurate problem in Blob analysis,the technique of jujube defect detection based on deep learning was proposed.Based on Inception_v3 model of GoogleNet,using Tensorflow,the artificial intelligence platform,setting and adjusting the learning rate,batch size and iteration number of the model.Then,the accuracy rate curve and losscurve of network training are obtained.The experimental results show that the model has good recognition performance on dried immature fruit,yellow skin fruit,skin crack fruit and mildewed fruit,the detection accuracy reach over 98%.The data-set collection and preprocessing is complicated,which leads to insufficient data-set samples and over-fitting in training steps.To solve this problem,the jujube defect detecition technology based on transfer learning is proposed.The related theoretical knowledge of transfer learning is studied,and the training method of freeze layer based on transfer learning is proposed,including pre-training of neural network,extraction of "bottleneck layer" features,and fine-tuning the parameters of the model.The influence of iteration times and the number of data-set on the accuracy of the model in transfer learning is studied,verified and compared the transfer learning method with the deep learning method.Transfer learning can achieve better training effects on fewer data sets,and can save a lot of training time.Finally,using C# and Halcon as software development platform,the camera shooting sequence and image processing flow were analyzed,and the jujube defect detection system was developed.