Image Classification Of Pests And Diseases Based On Saliency Detection

Agricultural production is the basic guarantee of people’s material life.Crop diseases and insect pests have a huge impact on the yield and quality of agricultural products.With the country’s agricultural modernization,the prevention and control of diseases and insect pests has attracted more and more attention.As an important part of pest control,the classification and identification of pests and diseases mainly rely on manual observation in the early stage,and the classification efficiency is low.However,the pest image classification technology based on computer vision can not only save a lot of labor costs,but also improve the accuracy of pest classification and realize pests.Intelligent prevention and treatment.Traditional pest image classification mainly relies on image processing technology and machine learning classification algorithms.It uses feature operators to extract image morphological and texture features and input them into the classifier to complete the pest identification task.With the emergence of deep learning in recent years,people have begun to use convolutional neural networks to extract deeper semantic information from pest images and classify pest images.In view of the complex background of the pest data set and the large differences between classes,this paper starts from the process of identifying objects by the human visual nervous system,and proposes a pest classification model Pest Net based on saliency detection.When the model classifies pest images,it is mainly divided into two processes:salient region location and feature fusion classification.In the salient region positioning stage,by introducing the attention mechanism,a bottom-up saliency detection module OPM is built on the basis of the full convolutional neural network.The overall model is a U-shaped structure and is deeply abstracted by fusing the pest images layer by layer.Semantic features and shallow spatial location features are used to convey the spatial location information of pests.The model outputs the predicted image of the pest area at each layer,and calculates the loss function with the pixel-level label,and guides the model to pay attention to the local details of the pest.This paper also designed a mixed multi-scale loss function,which can calculate the changes in the area position and morphological structure of pests at the same time,so that the saliency detection of the pest area is more accurate.In the feature fusion classification stage,Considering the dependence of fine-grained image classification on local detailed features,a multi-region and multi-dimensional feature fusion module MFFM is constructed.The module mainly performs two feature fusions.First,multi-mode bilinear pooling is used.In this way,the pest spatial location features provided by the OPM and the global abstract semantic features of the image provided by the basic model are merged to obtain the local detailed features of the pests;then the pests’ global semantic features,spatial location features,and local features are spliced together.The detailed features are fused together as the basis for fine-grained classification of pest images.In addition,this article also cleaned and labeled the pest data set,and visualized the pre-trained model.It was found that there was a phenomenon of deviating attention from the pest area during the model classification process,so a data enhancement method based on weakly supervised learning was designed.,Through the spatial location features provided by OPM,the saliency area of the pest is cut and masked,and the expanded image is re-input into the network for auxiliary training,guiding the model to pay attention to the detailed information of the area and improving the classification ability of the model.Train each model on the large-scale pests and diseases data set IP102.Firstly,the effectiveness of each module is verified by independent control experiments of the modules,and then compared with the training results of other pest classification models.The Pest Net model designed in this paper is on the IP102 pests and diseases data set.The accuracy rate of classification reached 70.13%,which is better than other models,and can accurately classify large-scale and multiple types of pests and diseases data sets.