Design Of Banana Disease Diagnosis System And Plant Number Detection Based On Deep Learning

Realizing the information management of banana plantation is one of the key means to improve the efficiency of banana production.The information management of the banana plantation mainly includes the automatic collection of growth information such as banana nutrients,plant diseases and insect pests,and the number of plants,the automatic recording of production materials invested in water,fertilizers,and medicines,and the automatic monitoring of the plantation environment.With the widespread use of imaging sensors on smartphones and drones,color digital(RGB)images have become one of the most accessible highthroughput data.In this paper,RGB images were used as the means to obtain the original information,and the deep learning method was used to study the common banana diseases and the early plant number detection methods.A remote diagnosis system for banana diseases and a banana plant number detection algorithm were designed.The main work of this paper are as follows:1.Construct a banana disease diagnosis model.A total of 5944 images were collected,including 7 common diseases of banana plants and healthy ones.The data set was randomly divided into training set,validation set and test set.The transfer learning method was used to train the deep convolutional neural network model Goog Le Net.By changing the model parameters and training times,the final diagnostic model was obtained.The final average test accuracy rate reaches 98% which has a good application prospect.2.Design of the remote diagnosis system.Further developed a remote diagnosis system using the mobile application of mobile phone(APP)as the terminal.The diagnosis model and middleware communication module were used as the server.Based on Android Studio,a simple mobile APP was designed.The APP can acquire banana images on-site and perform preprocessing.Then it can upload images to the remote server integrated with a diagnostic model via the Internet.Finally,the server receives the images,and then feeds the diagnostic results to the terminal APP.The system can quickly and effectively diagnose banana diseases online.3.Algorithm design of banana plant number detection model.A panoramic image of a banana plantation planted in a large area was obtained by aerial photography by drone,and the number of banana plants was detected using the Faster-RCNN target detection framework.In order to solve the problem of low detection accuracy of a single dense target in a large range,a "crop-recognize-splice" method was used to improve the accuracy of detection.The recognition model was obtained by training the manually labeled data set.The final F1 score is 94.99%.Combining the recognition model and the deduplication algorithm,the final average counting accuracy reaches 97.8%.