Research On Rice Pests' On-Line Identification And Directional Application Based On Improved YOLOv3

In the process of growing,Rice is easy to be attacked by pests,which threatens the yield and quality.In order to control the number of pests effectively,the on-line identification and directional application are studied in this paper.The research contents and experimental results are as follows:1.In view of the YOLOv3 network,disgning the improved DS-YOLOv3(Dense Net-SPPNet-YOLOv3)network by downsizing the backbone,adding improved Dense Net and Spatial Pyramid Pooling,expanding scales of the detection,and improving the way of selecting frames and so on,to enhance the ability of extractingof small targets' features.So that the model's speed and accuracy is increased.In addition,creating expriments to compare the new modle's performance on detecting of small targets,covered objects with YOLOv3,the experimental results show that: When the size of rice pests is less than 40pixel×80pixel,the DS-YOLOv3 network has obvious advantages over the YOLOv3 network.When the pest is blocking 50%?80%,DS-YOLOv3 can still maintain the ability to identify them,while the network identification accuracy of YOLOv3 is seriously reduced with the phenomenon of unrecognizing;The average detection accuracy of DS-YOLOv3 in the paddy field complex environment reached 82.3%,the average detection speed was 35 frames/s,the average detection accuracy of YOLOv3 increased by7.4%,the average detection speed of YOLOv3 increased by 9 frames/s,other detection items are also better than YOLOv3 network.2.Aiming at the problem of directional application,analyzing the positioning principle of binocular vision,and deducing coordinate transformation formula of binocular vision camera,the expiriment completed the calibration of binocular vision camera's interior and exterior parameters.Taking pests as identification objects,the positioning process and position parameters of the robot are calculated basing on binocular vision in the installation position of robot.3.Taking the embedded system Nano PC-T4 as the core master controller and combining with STM32 slave controller,the research and design of pesticide application robot' motion state and the control of he hot mist's flow valve state were carried out,and the software and hardware experimental platform of pest image collection,transmission and positioning was designed and developed based on the Nano PC-T4 development board.The image of the pest is transmitted to the remote server,which improves the real-time on-line identification.Meanwhile,the master-slave controller communicates with the serial port to realize the directional application of the pesticide application robot.4.The test results of the experimental platform show that the pesticide application robot can locate the rice pests through the pixel coordinates passed from the remote server.Aiming at the positioning error of binocular vision camera within plus or minus 1 degree of the experimental platform,20 sets of positioning 3d coordinate points were compared with the actual 3d coordinate points.The experimental results show that the angle error is within plus or minus one degree,and the distance error is within 20 mm.Through monitoring software program frame rate change always maintain in 19?32 frames/s meets the online real-time detection,and administer the robot's electrical and thermal fogmachine status bar has a data flow control valve of the motor feedback,hall sensor to detect pesticide robot speed curve changes,shows that pesticide application robot with the ability of independent orientation.