Research And Implementation Of Intelligent Monitoring Platform For Rice Damage Symptoms Of Pests On Rice Canopy Based On Deep Learning

The severe and frequent occurrence of rice pests and diseases have seriously affected the yield and quality of rice in China.The field survey of rice diseases and pests aims to obtain the incidence of pests and diseases,and provide a reliable basis for controlling rice diseases and pests.At present,the methods of manual field surveys have some problems,such as high labor intensity,strong subjectivity,and untraceable survey results.In order to realize the intelligent field investigation of rice diseases and pests,this paper designs and implements an intelligent monitoring platform for pests on rice canopy,which can detect rice damage symptoms of pests on rice canopy in real time.The main research contents and results are as follows:(1)Research and improvement of rice canopy pest detection algorithm based on deep learning.As the damage areas of rice leaf roller and cockroaches on rice canopy occupy a small area in the monitoring image,the location and size are variable.In order to improve the detection effect of the two pests as damage areas,Faster R-CNN detection model based on Region Proposal and RetinaNet detection model based on regression are studied and improved.Faster R-CNN detection model based on Region Proposal and RetinaNet detection model based on regression.For the improved Faster R-CNN detection model,in order to improve the overall performance of the network,candidate area generation network and classification regression network are separately set up as feature extraction network,respectively.Multi-scale RPN network structure based on three different size sliding windows is adopted to improve the detection effect of small targets.For the improved RetinaNet model,Resnext-101 is used as the feature extraction network in the improved RetineaNet model,and the feature pyramid network structure is improved to improve the detection performance of small targets.The Group Normalization method is used to solve the problem of high error rate caused by small batch Normalization.Focal loss was chosen as the loss function to solve the problem of the serious imbalance between positive and negative sample ratios in a single-phase detection network.(2)Training,testing and analysis of rice canopy pests as a pest detection model.To improve the generalization ability of the model,the original image is enhanced by mirror flipping,increasing brightness and contrast,and adding Gaussian noise.In order to verify the effectiveness of the improved algorithm in this paper,transfer learning methods are used to train the models before and after the improvement.In addition,the AP,mAP,FPS and P-R curves of each model are compared under the same test set.The results show that mAP based on the improved Faster R-CNN is 7.32%higher than the original algorithm,and mAP based on the improved RetinaNet is 5.72%higher than the original algorithm.To guarantee the real-time and accuracy of the platform,the improved RetineaNet detection algorithm is selected.Among them,the average detection precision of rice canopy pests was 93.76%and the detection rate was 12 frames/second.(3)An intelligent monitoring platform for rice canopy pests was designed.The hardware of the platform includes network cameras,equipment boxes,gimbals,servers and PCs.In this paper,the platform software was designed,which can control the collecting equipment to collect the images of rice canopy pests and send them back to the server in real time.Then,the detection results of the pest image are displayed and saved locally.In this paper,an intelligent monitoring platform for rice pests based on object detection algorithm is presented.In the actual monitoring scenario,functions such as detection and early warning and result storage can be realized to meet actual needs.