Research And Optimization On Rice Light-trap Pest Detection Method Based On Deep Learning

In China,the loss of rice caused by pests is very serious every year.real-time and accurate monitoring of rice pests is an important premise to reduce economic loss.It is easy to monitor rice pests in the field by means of measuring and reporting lamp,and it can trap many kinds of pests,so it is widely used.At present,it is still necessary to identify and count target pests manually in the images collected by the measuring lamp,which is time-consuming,laborious and subjective.With the development of artificial intelligence,it is possible to recognize and count pests automatically by machine vision technology.In this paper,an improved YOLO-pest detection model,YOLO-pest detection model based on Yolo-v3,was proposed to solve the problem of false detection and missed detection caused by insect adhesion in the light-induced insect images.Three pests(Cnaphalocrocis medinalis,Chilo suppressalis and Sesamia inferens)were automatically detected and counted by YOLO-pest model.The main results are as follows.(1)Research on different detection algorithms of rice light-trap pests based on deep learning.Target detection algorithms are generally divided into two-stage target detection algorithms based on region nomination(such as Faster R-CNN,Cascade R-CNN)and one-stage target detection algorithm based on end-to-end(such as Yolo series,SSD).In this paper,two pest detection models based on Faster R-CNN and Yolo-v3 are trained respectively.The results show that the model based on Yolo-v3 has higher precision.(2)Research and optimization on detection algorithm of rice light-trap pests based on YOLO-pest.As the backbone network of YOLO-v3 model Darknet-53 lost some feature information in the process of convolution and down sampling,and the original nine-scale anchor was not suitable for three rice pests.In this paper,we improve the YOLO-v3 model,and adopt DenseNet instead of the low-resolution transport layer of Darknet53 in the model.The feature extraction network in YOLO-pest is helpful to feature propagation,promote feature reuse and network performance improvement.The improved k-means algorithm is used to calculate the anchor boxes of rice light trapping pest data set,which made anchor fit the size of rice light-trap pest better and improved the detection rate of bounding box.(3)Training,testing and analysis of rice light-trap pest detection model based on deep learning.Two Faster R-CNN models based on ResNet-101 and VGG-16 and YOLO-v3 model and YOLO-pest model based on Darknet-53 are trained for the same data set.The results of the same test set show that the YOLO-pest is the best,the mAP is 94.9%,and the annual image of a site is tested.The results show that the average accuracy of the three target pests is 96.3%,the average correlation coefficient between the result of automatic detection and the result of artificial identification was 0.93.The YOLO-pest detection model proposed in this paper can automatically detect three species of rice light-trap pests(Cnaphalocrocis medinalis,Chilo suppressalis and Sesamia inferens),and reduces the false detection and missed detection caused by the touching insects in light-trap images.The automatic detection data has a high correlation with the manual identification data.The automatic detection method can be used for intelligent monitoring of rice light-trap pests.