Study On Lightweight Recognition Method For Wheat Pests

Wheat is the second largest food crop in China,so ensuring its yield is crucial to food security.Wheat diseases and insect pests will cause wheat yield reduction and affect food security.With the development of AI(Artificial Intelligence)and UAV(Unmanned Aerial Vehicle)measurement technology,the wheat pests recognition can be realized through UAV inspection and deep learning,and the area of wheat pests can be found by combining GNSS(Global Navigation Satellite System)and GIS(Geographic Information System)to prevent pests and diseases.However,the mainstream wheat pest recognition method based on convolutional neural network is deployed on the central server.The real-time performance of image data transmission is easily affected by the network bandwidth between the UAV and the server,and the server performance requirements are also high.Therefore,studying the lightweight wheat pest recognition method to achieve real-time object recognition is the key to solving this problem.At present,with the development of edge computing,some scholars have studied the method of lightweight deep learning model,moving computing tasks to network edge devices,alleviating the pressure of central servers and the delay of image data transmission.But in the face of small targets,multi-scale,dense and mutually occluded wheat pests,the lightweight model obtained by this method still has problems such as loss of pest feature information,inaccurate positioning of boundary frames,and missed recognition.Aiming at these problems,this thesis studies the improvement of the existing lightweight recognition method to make it suitable for wheat pest recognition.The main research results of this thesis are as follows:(1)Aiming at the problem that the YOLO V5 lightweight wheat pest recognition method based on slimming pruning fine-tuning has low recognition accuracy due to the loss of wheat pest feature information,a lightweight wheat pest recognition method based on knowledge distillation with local enhancing is proposed.Firstly,the pest and background features are enhanced by spatial and channel attention,and the mask is used to separate the pest features from the background features for local distillation.Secondly,the joint knowledge distillation strategy is used to separate the knowledge,and then the feature knowledge extracted by the heavy wheat pest identification model is distilled to the lightweight wheat pest recognition model.Finally,the proposed method is used to train the lightweight model on the wheat pest training set.The optimal lightweight wheat pest recognition model is verified,analyzed on the wheat pest verification set,and compared with other lightweight wheat pest recognition methods based on model compression.(2)Aiming at the problem that the Yolo-FastestV2 lightweight wheat pest recognition method based on manual design identifies small,dense and mutually occluded pests with inaccurate bounding box positioning and missed recognition,a lightweight wheat pest recognition method fusing location enhancement and adaptive label assignment is proposed.This method first improves the original LightFPN structure,and fully extracts the pest location information and semantic information by fusing multi-scale feature maps to enhance the bounding box positioning ability.Secondly,the adaptive anchor box clustering is combined with the adaptive label assignment,and the anchor box with similar shape and the largest Io U is assigned to each pest to be identified,so as to avoid the missing allocation of anchor boxes for a pest to be identified,thus alleviating the phenomenon of missing pests due to dense and mutual occlusion.Finally,the model is trained on the wheat pest training set.After training,the optimal lightweight wheat pest recognition model is obtained to verify and analyze on the verification set.The recognition effect is tested on the test set and compared with other lightweight wheat pest recognition methods based on artificial design.(3)In order to use UAV edge computing to detect wheat pests,a lightweight framework of wheat pest recognition for UAV edge computing is constructed by combining the theoretical knowledge of the two methods proposed.The feature knowledge extracted from the wheat pest recognition model based on YOLO V5 is distilled to the distillation framework of the lightweight wheat pest recognition model that combines location enhancement and adaptive label assignment.The framework is used to train the lightweight model.The lightweight model is embedded in the edge equipment to detect wheat pests,which proves the correctness and practicability of the proposed method.