Apple Foliage Diseases Recognition In Android System With Transfer Learning-based

Apple is one of the main fruits in our country,which has brought rich benefits to our economy.However,the quality of apple is susceptible to different diseases,so the recognition and classification of apple diseases is particularly important.This paper takes five common apple foliage diseases in Loess Plateau as research objects,such as alternaria leaf spot,brown spot,mosaic,gray spot and rust,and proposes an Android detection system with transfer learning for five apple foliage diseases.The main work of this paper is summarized as follows.(1)On the basis of the common characteristics analysis of apple foliage diseases,to prevent the over-fitting problem of network training caused by insufficient data samples,this paper uses data enhancement to increase the diversity of training samples.Through eleven operations including image rotation,horizontal and vertical mirroring,sharping,brighting,contrast adjustment,and gaussian blurring,a total of 2029 apple foliage diseases data set of 5 categories are increased to 24,348 after manual screening.To obtain standard format for the subsequent network model training,the data set has been made into VOC2007 format.(2)On analysis basis of parameters transfer learning,to improve the efficiency of network model training and to obtain better diseases detection effect,this paper proposes a parameter transfer model on ImageNet data set,where the initial parameters of the model are obtained by pre-training the VGG-16 and ZF Net on the ImageNet data set,and fine-tuning training transferred model conducted by the data set of apple foliage diseases.The Faster RCNN algorithm is employed to detect apple foliage diseases includes three parts of feature extraction network,the candidate regions generate networks and classification regression network.The feature image of the disease image data set can be obtained by VGG-16;the possible candidate regions of the disease can be achieved by the candidate regions generate networks of RPN,and the final detection areas and results of five kinds of diseases can be obtained by its classification regression network.Comparisons with Fast R-CNN algorithm using VGG-16 as feature extraction network and Faster R-CNN algorithm using ZF network as feature extraction network,the validation effect of this paper's proposal is better and the AP(Average Precision)shows as follows: rust detection effect is the best of 90.35%,and the mosaic 81.66%,alternaria leaf spot 71.64%,brown spot 75.67%,and the effect of gray spot is the worst of 63.44%.Five kinds of diseases of the mAP(mean AP)value is about 76.55%.(3)To apply the training result of apple foliage disease based on Faster R-CNN algorithm into the Android system,this paper designs and implements an Android detection system.Firstly,on the basis of analyzing the requirements of the Android detection system,the training model of Faster R-CNN is transplanted to the Android platform,and the Android environment is configured in three aspects.Secondly,an Android client with two functional modules is designed and developed;Finally,the Android system has been tested with 200 trial samples,which are randomly selected 40 images of each apple foliage disease.The experimental results show that the Android system can meet the requirements of recognition accuracy,and the corresponding detection accuracy lists as follows: the detection accuracy of rust can be up to 100%,the detection accuracy of alternaria leaf spot and mosaic up to 97.5%,the brown spot up to 95%,the gray spot up to 92.5%,and the comprehensive detection effect up to 98.5%.