Research And Implementation Of Small Object Detection Method For Wheat Fusarium Head Disease Based On Complex Backgroud

Crop diseases are one of the inevitable problems in agricultural production,which can cause serious yield loss and quality reduction,and even endanger the growth and development of crops and life safety.As one of the three major edible grains in China,wheat is vulnerable to various diseases throughout the growth process,among which fusarium head blight occur most seriously in the wheat areas south of the Huaihe River and in the middle and lower reaches of the Yangtze River.The main symptoms of red mold are red or orange spots on the surface of wheat seeds,and the production of mycotoxins.Therefore,how to effectively prevent and control wheat erysipelas to protect the crop is an important issue for the development of agriculture in China.In order to meet the needs of the development of smart agriculture and solve the problems of low efficiency and poor reliability of wheat fusarium head blight detection,this paper takes wheat fusarium head blight as the research object from practical applications,combines the algorithms of YOLOv5 and Transformer in deep learning to achieve the detection of small object of wheat fusarium head blight in complex backgrounds,and designs a Web-based wheat fusarium head blight detection system.In this paper,the following three main aspects of research were conducted:(1)A complex background detection algorithm based on YOLOv5 combined with Swin Transformer is proposed,which effectively solves the problem of disease identification by complex background interference.For the problem of complex background interference and relatively blurred dense object when extracting the disease features of wheat fusarium head blight in large fields,the ESRGAN super-resolution model is introduced at the input of YOLOv5 to highlight the features of the inspected object and mitigate the influence of background noise.Also taking advantage of Swin Transformer and YOLOv5,Swin Transformer is introduced into the YOLOv5 algorithm as the backbone network of YOLOv5,which enables the interaction of adjacent windows to enhance the global modeling capability due to the presence of the shift-windows multi-head self attention modules(SW-MSA)mechanism.This integration can compensate for the lack of YOLOv5 as a typical CNN(Convolutional Neural Network)network capturing global and contextual information to improve recognition accuracy.Finally,the output prediction layer uses the Soft-NMS function instead of the NMS function to filter the prediction frames to reduce the object miss rate in dense areas.Ablation experiments and comparison experiments were conducted on a homemade wheat blast dataset to verify the effectiveness of the improved strategy and the superiority of the proposed method in this chapter.The experimental results showed that the mAP(mean Average Precision)was effectively improved by 3.5%to 94.4%when compared with the YOLOv5 algorithm.Therefore,the improved Sw-YOLOv5 algorithm is able to detect fusarium head blight effectively and provide technical support for fusarium head blight detection in different complex environments.(2)A small object detection algorithm with complex background is improved to effectively solve the problem of small object false detection and missed detection.Based on the above complex background detection algorithm,firstly,the C3NRT structure is introduced in Neck instead of the original C3 module to obtain global information to improve the accuracy of small object detection and reduce the computational complexity.Secondly,the ECA attention mechanism is adopted to fuse to the effective features after the feature pyramid,capture the channel correlation of CNN,and achieve the channel adaptive calibration of the feature map,so as to increase the attention to the object region.Then,a layer is added to the original detection layer and an additional feature fusion structure is added to the Neck accordingly to enhance the algorithmic model to detect small-scale object information.Finally,the convergence speed of the network is accelerated and the accuracy of the model detection is improved by fusing the decoupled head in the detection head and separating the classification and regression operations.Ablation experiments are conducted for each improved module,and the experimental results show that the improved small target detection algorithm has significantly improved the detection accuracy of small targets compared with the original algorithm,and the multiple improvement strategies improve the mAP by 3.1%.The proposed algorithm can provide farmers with more accurate early disease detection,reduce the probability of disease occurrence,and improve their economic benefits.(3)A Web-based wheat fusarium head blight detection system was designed to achieve wheat fusarium head blight detection.Based on the above method,the wheat fusarium head blight detection system is designed.It consists of three parts:client side,server side and database.The improved model is deployed to the server side and the client side is presented in the form of a web page to provide a visualized wheat fusarium head blight detection platform for agricultural workers and researchers.When users conduct disease detection,the server side uses the constructed disease detection model as the core component and the MySQL database as the data support to complete the tasks of disease detection and data storage to meet the actual needs of agricultural production.