Research On Identification Technology Of Major Corn Pests Based On Deep Learning

Food production is a matter of livelihood.As the second most important food crop in China after rice,the yield and quality of corn is immeasurably damaged by corn pests.The prerequisite for effective pest management is the rapid and accurate identification and detection of pests.The efficiency of traditional detection by manual visual observation is too slow and requires too much experience and knowledge of staff.And the operation and maintenance costs of professional detection equipment are too high in the context of decentralization of farmland in China.All these do not meet the needs of farmers for real-time detection of corn pests.In order to achieve more efficient and convenient detection,the thesis is based on deep learning for corn pest identification technology.The main research work is as follows:(1)In order to meet the demand of efficient and convenient real-time detection,the paper proposes a lightweight target detection algorithm to improve YOLOv5,starting from convolutional neural networks and using the One-stage YOLO series target detection algorithm as the main technical route.First,the lightweight network PP-LCNet is used as the backbone.Complete the research and adjustment of depthwise separable convolution,activation function and SE attention mechanism modules to reduce the FLOPs and parameters in the network inference process.Secondly,the Add layer of weighted feature fusion is proposed to enhance the learning of the underlying feature information of the pest,so as to improve the detection of complex backgrounds and small-sized targets.Finally,the loss function SIo U that considers the angular relationship between prediction frames is introduced to improve the regression efficiency of prediction frame.(2)In view of the lack of corn pest datasets,the main corn pests are selected from domestic and foreign public pest datasets.Complete some pest categories with insufficient data by manual supplementation.Distinguish and categorize the different life cycle forms of the same pest.Thus,a corn pest dataset with a total of 5302 original images in 8 categories was constructed.Based on this dataset,data enhancement was performed by YOLOv5,and the ablation experiments for the above improvements were completed under the same environmental parameters.The comparison results show that the lightweight corn pest detection model with improved YOLOv5 ensures the accuracy of identification detection while significantly reducing the model complexity.(3)Based on the improved detection model,the tflite porting of the model is completed in the Android platform,and complete the development of the real-time detection application for the images acquired by the camera.The test results show that the proposed method in the thesis can meet the requirements of recognition accuracy and detection speed in agricultural work,and the deployed application runs fast in mobile.