Research On Target Detection Technology Of Plant Protection UAV Based On Deep Learning

Crop diseases and insect pests are the two most important factors leading to the reduction of food production in the world,although the application of plant protection UAV can automate pesticide spraying,the identification and detection of crop diseases and insect pests still need a lot of manpower and time cost.Deep learning technology can solve this problem,the use of plant protection UAV for the detection of crop diseases and insect pests can achieve intelligent and refined pest control.With the continuous development of computer hardware and deep learning theory,the object detection algorithm based on deep learning has become more and more mature,but the current mainstream object detection based on deep learning algorithm is difficult to be directly applied to plant protection UAV.On the one hand,the targets of crop diseases and insect pests are complex and variable in scale,so the detection accuracy is generally very low.On the other hand,speed is a requirement for the detection of crop diseases and insect pests based on protection UAV.However,improving the detection accuracy will sacrifice the detection speed to a certain extent,which can’t meet the real-time requirements.Therefore,it is very important to research a real-time object detection algorithm with high detection accuracy.This paper makes an in-depth research on the above difficult problems,and proposes an object detection algorithm SepRes-YOLOv3-Tiny which can be used in plant protection UAV based on YOLOv3-Tiny algorithm.First of all,this paper designs a new convolution module based on depth separable convolution and residual structure to replace the traditional convolution to extract image features to improve the detection accuracy.Secondly,the detection layer network of YOLOv3-Tiny algorithm is increased to three scales.This multiscale detection based on feature fusion can improve the recall of the algorithm.Finally,the GIOU is introduced for bounding box regression and a priori anchor box clustering to further improve the recall of the algorithm and the positioning accuracy of the object boundary box.After analyzing the feasibility and advantages of this algorithm in theory,this paper also designs five comparative experiments on IP102 and Plant Doc data sets.First of all,three experiments are carried out to verify the effect of the improvement of this algorithm in detail.Secondly,one experiment is carried out to compare and analyze the algorithm in this paper and YOLOv3-Tiny algorithm as a whole.Finally,the algorithm in this paper are compared with other advanced algorithms.By analyzing the values of various evaluation indicators,the experimental results show that the algorithm has higher detection accuracy and can basically meet the real-time performance of object detection.The object detection algorithm designed in this paper can meet the requirements in terms of detection accuracy and speed,and can be applied to plant protection UAV to detect crop diseases and insect pests.Therefore,the algorithm proposed in this paper has very important practical application value for intelligent and automatic agriculture in the future.