Study On Identification Of Crop Diseases And Pests Using Lightweight Convolutional Neural Network

At present,the new generation of information technology represented by artificial intelligence has infiltrated various fields.With its outstanding performance in computer vision,the relevant theories and applications of CNNs have been rapidly developed,laying a good foundation for intelligent and accurate identification of crop diseases and pests.Generally,the recognition accuracy of CNN is proportional to its structural complexity.While the existing research has achieved good recognition results,the model size is generally too large and the computation is too high,which is not conducive to mobile terminal applications in the edge computing environment.Therefore,this article intends to focus on the identification of crop diseases and pests,focusing on the establishment of lightweight CNN models and their recognition methods.Compared with methods that use complex structures to improve recognition accuracy,the research difficulty of lightweight model recognition methods lies in ensuring high recognition accuracy.Therefore,this paper takes Plant Village,AI Challenger 2018 and self built specific pest datasets,which are widely used in disease classification research,as data sources,respectively,based on typical CNN SqueezeNet,SqueezeNext,MobileNetv2-SSDLite,studies and establishes lightweight models,and proposes crop disease and pest identification methods based on lightweight models to facilitate the application of mobile terminals.The research content of this article is as follows:(1)A lightweight CNN model of SqueezeNet is established,and a high accuracy multi class leaf disease classification method based on the lightweight model is proposed.In response to the shortcomings of large model size and high computational complexity in existing multi category leaf disease classification,combined with the classification task of 38 categories on the Plant Village leaf disease dataset,on the one hand,by reducing network depth and adjusting the proportion ratio of 1×1 and 3×3 convolution kernels in the expansion layer of fire module,reasonable postposition of some fire modules,to design lightweight reconfiguration of the SqueezeNet backbone network,on the other hand,by using feature fusion between the outputs of the fire module and combining softmax loss and center loss to improve the recognition accuracy.The experimental results show that the accuracy is 98.47%,the model size is reduced from 2.91 MB in the original SqueezeNet to 1.68 MB,and the floating-point operation amount is reduced from 272 MFLOPs to 145 MFLOPs.This indicates the effectiveness of this method for high accuracy multi category leaf disease classification.(2)A lightweight CNN model of SqueezeNext is established,and a high accuracy fine grain classification method for multi class leaf diseases based on the lightweight model is proposed.In view of the deficiency that the existing fine grain classification of leaf disease mostly adopts complex structure,and in order to meet the requirements of high-performance and miniaturization of embedded applications,in combination with the fine grain classification task of 59 categories on the reconstructed AI Challenger 2018 leaf disease dataset,on the one hand,the SqueezeNext backbone network is reconstructed and designed by integrating multi-scale feature extraction and coordinate attention mechanism,on the other hand,by using FPN feature fusion and combining softmax loss and modified central loss to improve the accuracy of fine-grained classification.The experimental results show that the accuracy reaches 92.69%,which is 3.77 percentage points higher than original model and 1.71 and 1.07 percentage points better than Res Net50 and Xception.The model size is only 3.86 MB,which is only4.2% and 4.7% of Res Net50 and Xception.The model accuracy density is 23.89%,which is 24.4 and 21.5 times higher than the two,respectively.This indicates the effectiveness of this fine-grained classification method in improving recognition performance and efficiency.(3)The MobileNetv2-SSDLite lightweight CNN model is established,and a high precision pest target detection method based on the lightweight model is proposed.Targeting specific pest target detection in research on disease and pest identification,combined with the target detection task of nighttime cicada nymphs,based on the target detection network MobileNetv2-SSDLite,using CA＿MobileNetv2-SSDLite as a teacher model,appropriately cuts its depth and width as a student model,uses the knowledge distillation algorithm based on attention transfer to train the student model,and uses teacher model knowledge to guide the student model from three aspects of feature extraction,target classification,and border prediction in the construction of the target detection loss function.The experimental results showed that the model size was significantly reduced from 13.7MB to 3.88 MB,and the floating-point computational complexity was significantly reduced from 667.73 MFLOPs to 196.24 MFLOPs.The average precision of object detection reached 91.64%,which is roughly equivalent to the current mainstream lightweight models YOLOv4-Tiny and YOLOX-Tiny.However,the model size is 17.3% and 20.0% respectively,and the detection speed is more than twice that of two.This indicates the rapid and effective detection of high precision pest targets using this object detection method.(4)Apply the fine-grained classification method for leaf diseases proposed in this article to the classification of 6 types of outdoor apple leaf diseases.At the same time,apply the pest target detection method proposed in this article to the outdoor insect target detection composed of 3 types of outdoor flying insects.The experimental results indicate that the lightweight leaf disease fine-grained classification method proposed in this article can also be applied to the disease classification of outdoor apple leaves.The lightweight pest target detection method proposed in this article can also be extended to the target detection of other outdoor insects.This article focuses on the problem of disease image classification and pest target detection in disease and pest recognition,and conducts research on recognition methods based on lightweight models,laying a theoretical and technical foundation for the mobile application of disease and pest recognition.