Identification Of Citrus Diseases And Insect Pests Based On SSD Algorithm

With the continuous development of deep learning,its application in life is more and more extensive.In this paper,the SSD target detection algorithm in deep learning is mainly applied to the identification of pest and disease leaves and fruits in citrus production process,and the pretreatment of image data is improved to solve the problem of small sample number and serious misclassification,so as to improve the accuracy of the model.In this paper,648 pictures of 19 common diseases and insect pests in citrus planting were collected,and label Img software was used to label each picture.Then,according to SSD algorithm,combined with Python programming,a preliminary model of citrus pests and diseases target detection was realized.AP value was used to evaluate the recognition effect in the verification process.The results showed that some classes had higher AP value,but the recognition effect was not ideal.In this paper,the preliminary model is improved in five steps.The first step is to improve the model evaluation criteria,and adopt annotation number,FN,FP and TP criteria to evaluate sample size,missed classification,misclassification and correct classification respectively.Step 2: For each image sample output of verification and test,four result images of misclassification,missing classification,machine recognition and manual annotation are combined into one image,so that possible problems of each sample can be intuitively found.The third step is to increase the sample size by data enhancement because the sample size of some classes is too small,resulting in over-fitting.The fourth step is that the preliminary model is easy to misidentify the interference pictures of non-orchard background.This paper adopts the interference pictures as the background,pastes the screenshots of the leaves or fruits of pests and diseases,manually marks the pictures,and adds into the original samples as samples for training,verification and testing.The fifth step is to adopt the flask network framework deployment model and display it in user interface.After the improvement of the data in steps 3 and 4,the mean value of FP/ label representing misclassification decreased from 785.69% to 14.66%.The feature of this paper lies in the perfection of model evaluation criteria and the treatment of interference image misclassification.The significance of this paper is that deep learning can be applied to agricultural citrus production,which can not only reduce labor and cost in the production process for farmers,but also bring high quality fruit consumption to the people.