Study On The Method Of Tumor Root Classification Based On Convolution Neural Network

Rape occupies an important position in my country’s oil crops.It has many excellent characteristics such as high yield and high quality,so it can bring very high economic benefits and practical value.The disease can lead to the development of clubroot disease in healthy canola roots,which will greatly reduce the yield and may even directly lead to the death of the rapeseed.If the knowledge of plant phenotypes is used to accurately detect the disease degree of canola clubroot and achieve accurate classification of canola clubroot,it is possible to accurately evaluate the resistance of rapeseed varieties,and promote the selection and breeding of excellent varieties by professionals.The study of defense against rapeseed diseases has far-reaching significance in many aspects.At this stage,the method of grading rapeseed mainly relies on the manual grading of experienced professionals,and the grading standard also completely depends on the subjective judgment of the grading personnel,which leads to the whole grading process being very time-consuming and inefficient.At the same time,there are still problems such as high error rate and strong subjectivity,and this method will not be applicable if faced with the huge amount of image data of rape roots.With the rapid development of artificial intelligence convolutional neural network,combined with image processing and embedded technology,the image data of rape root tumor is trained as a grading model of root tumor,and the model is transplanted into embedded equipment to design a fully automated the developed rape root tumor grading system can greatly improve the grading efficiency and accuracy of rape clubroot.This paper mainly starts with traditional image processing,performs image segmentation and feature extraction on the rapeseed root in the original image,highlights the tumor characteristics of the original rapeseed root image,and then performs convolution training on the image to obtain a rapeseed root tumor grading model,and compares different The method of training the model is the most efficient way of training,and finally the model is transplanted to the embedded device to complete a system that can automate the grading of rape root tumors.This paper mainly completes the following research work:(1)Image preprocessing.There is a lot of noise information in the root image of rapeseed,so it is necessary to perform image preprocessing on the image to highlight the character information of the root tumor,so that a model with high accuracy can be obtained during convolution training.The specific method is to first crop the original image,crop the image of multiple rapeseed plants into a single plant,and then use the image segmentation method to distinguish the target root region image from the background image,and filter out the background noise data.(2)Rape root tumor data set and establishment of grading model.The rapeseed root tumor original image and grading standard provided by the Oil Crops Research Institute of the Chinese Academy of Agricultural Sciences will be preprocessed by image segmentation to create a dataset of 5000 rapeseed root images through Label Img.The BP neural network is then used to train the grading model of canola clubroot.In order to obtain the best grading model,a variety of model structures were selected for comparative experiments,and the grading model of canola clubroot was obtained through GPU training.The parameters are adjusted to optimize the model.Through the experimental data,it is found that the Res Net-50 network structure has the highest accuracy rate,and the accuracy rate of using the test set test is 92.00%.(3)Use TPU for model training and optimize the model.Although the classification accuracy rate of the classification model trained by GPU can reach more than 90%,it takes up a lot of computer resources in the process of training the model,the training efficiency is low,and it is not conducive to transplanting the model to the embedded device,so it needs to use TPU for model training.The Caffe convolutional network training environment was built in Ubuntu 18.04 to complete the construction,optimization and training of the rape root tumor model,and rely on API functions to call the TPU processor to perform deep neural network computing tasks.Through the experimental data,it is found that the TPU training model has considerable advantages in terms of computer resource occupation,power consumption of voltage and current,and calculation speed without affecting the accuracy.It is also more convenient to transplant to embedded type equipment.(4)Construction of rape root tumor grading system.Transplant the model to the ETPU-Z2 development board,build a cross-compilation environment through Linux,transplant the Open CV dependency library and the root tumor grading model trained by TPU,build and debug the software and hardware,and form a complete grading system together.