Research On Detection Of Wheat Mites In Wheat Fields Based On Deep Learning

Wheat is one of the important food crops in China.Increasing the yield of wheat will help improve national economy and people’s living standards.In recent years,there is an aggravating tendency in wheat pests outbreak year by year,and speedy and accurate access to disease dynamics statistics has become an urgent need.As wheat mite,one of the wheat mites,is small size,it is still visual estimate in the field.In order to improve the timeliness and accuracy of monitoring population dynamics of wheat mites in the field,a large number of target detection methods in computer vision technology have been successfully applied to the detection of crop pests.These methods have the advantages of faster,more accurate,objective and convenient,and at the same time greatly promote agricultural modernization.Compared with the pest images obtained under the experimental conditions,the pest images in the field are more difficult to detection due to the complex farmland environment,unstable illumination,and different postures.Traditional machine learning target detection algorithm based on sliding window region selection strategy has the disadvantage of untargeted,high time complexity,window redundancy,and artificial design characteristics,and has no good robustness for diversity change.With a certain degree of invariance to geometry,illumination and deformation,deep learning has the ability to adaptively construct feature descriptions,possessing better flexibility and generalization capabilities.In this paper,the automatic identification and counting of wheat mites in the field is studied by using the deep learning target detection algorithm.The main research contents and innovation points are as follows.1.The models of pedestrian detection based on convolutional neural networks have been studied in recent years.Two models are mainly introduced.One is based on candidate region extraction models such as RCNN.Fast-RCNN and Faster-RCNN,and the other is based on the Regression models,such as YOLO and SSD.Through the analysis and comparison of the experimental results of wheat spider images,it is found that the region-based model is more suitable for the identification and counting of wheat mites in the field,and high recognition results can be achieved by improving the structure and important parameters of the network.2.The Faster R-CNN algorithm was proposed for detection of wheat mites in the field,and its performance was verified by experiments.This paper also improves the accuracy of wheat mites detection by adding two inceptions module on the basis of ZF Net for feature extraction and designing new anchor scheme,and verifies it through experiments,and the mAP is 0.8961.3.An improved algorithm based on Faster R-CNN isintroduced to deal with the problem of poor recognition effect of Faster R-CNN algorithm on the original image(resolution 1440*1080).The following improvements have been made to datasets and algorithms:on the dataset,the training set was made into four kinds of small images of different scales for training,and mirror image flip,salt and pepper noise were performed,and in the algorithm,the scientific design of anchor is carried out by using clustering algorithm,The position sensitive score graph(Positive-sensitive score maps)in the R-FCN algorithm is introduced and the important parameters in the network are optimized.Through the test verification,greatly increased the recognition effect to the original image,and increased the training speed.Through multi-scale processing of the dataset,the experiment shows that the mAP is improved by about 1%,and the robustness of the model is increased,which can have a good recognition effect on the original image.