Research And System Implementation Of A Lightweight Model For Identification Of Larval Pests In The Field

Pests are one of the main causes of damage to crop yield and quality.Accurate and fast identification and diagnosis of pests helps to control them timely and effectively,especially for early identification of pests in the larval stage.In recent years,image recognition based on deep learning has been increasingly used in pest image detection and recognition tasks,however,the large number of pest species in the field,interspecies similarity and complex backgrounds pose a great challenge for the majority of agricultural producers and monitors to identify pests and obtain useful information,while not being able to monitor pests in real time.Therefore,this study takes pest larvae as the research object from the need of mobile identification diagnosis in the field,and explores the light-weight recognition model of pest images in the field and its implementation method:(1)Construction of dataset: Pests were photographed using DSLR cameras in the surrounding areas of Beijing,and a total of 30 categories of pests were collected,with a total of 2862 images,and the dataset was expanded through a random,pipeline-based data enhancement method to construct the final pest dataset.(2)Base model comparison and selection: Through the comparison of three classical deep neural network models,four lightweight neural network structure characteristics,and the performance evaluation results on ImageNet and pest dataset,EfficientNetV2 is selected as the base model for this research.(3)Model lightweight method research: a new pest classification method PCNet(Pest Classification Network)based on lightweight CNN embedded attention mechanism is proposed.PCNet adopts EfficientNetV2 as the backbone network and introduces coordinate attention mechanism(CA)to learn in-channel pest information of input image and pest location information of the input image.In addition,the feature fusion module is developed by combining the feature mapping of the moving backward bottleneck(MBConv)output and the feature mapping of the averaging pooling output to achieve feature fusion between shallow and deep layers to solve the problem of insect pest feature loss in the downsampling process.Experimental results show that the PCNet model achieves 98.4% recognition accuracy on a self-built dataset consisting of30 classes of larvae,outperforming three classical CNN models(AlexNet,VGG16,and ResNet101)and four lightweight CNN models(ShuffleNetV2,MobileNetV3,EfficientNetV1 and V2).To further verify its robustness on different datasets,this study also tested the model on two public datasets,IP102 and mini ImageNet.pcNet outperformed the other models with 73.7% recognition accuracy on the IP102 dataset and 94.0% on the mini ImageNet dataset,which was only lower than ResNet101 and MobileNetV3.The number of parameters of PCNet is 20.7M,which is less than the traditional classical CNN model.The high accuracy and small size of the model make it suitable for real-time pest identification on resource-constrained mobile devices.(4)Mobile prototype system design: A pest larvae identification application was developed on the Android platform,which contains three main functional modules.The first module implements the user registration and login function;the second module adopts intelligent recognition technology to realize the recognition function of pest images,users can use the built-in camera of cell phones to take photos or upload the images to be recognized through photo albums,and crop them and upload them to get the recognition results;the third module implements the user labeling function of pests,users can take photos or upload images through cell phone photo albums,and The third module implements the pest user labeling function.After the labeling is completed,the user selects the category of the pest and publishes the name of the pest he/she has judged.Overall,this study proposes a new lightweight pest larvae classification method based on deep learning using pest larvae as the research object,and demonstrates the effectiveness of PCNet in achieving high accuracy at small model scale.The developed Android application can provide an easy-to-use platform for users to identify and annotate pests in real time.