| Brown planthopper(BPH)is one of the main rice pests,which affects rice yield to some extent every year.When investigating the number of brown planthopper,agricultural plant protection personnel commonly use the bat-bat method,visual inspection method and sweeping net method,etc.These methods require the investigators to identify and count brown planthopper with the naked eye.When the investigation takes a long time,the investigators tend to feel tired,which leads to the decrease of the accuracy of the investigation.In recent years,deep learning technology has developed rapidly and achieved many outstanding achievements.Researchers began to apply deep learning technology in their own fields.This project mainly studied the use of depth target detection network to detect and count brown planthoppers in local images of rice.The main research contents and achievements of this topic are:(1)A number of brown planthopper images were collected in the experiment,and the clearer images were selected and made into a data set in VOC2007 format after preprocessing,which was used to train the depth target detection network.Firstly,the brown planthopper image detection algorithm based on Faster R-CNN was studied.Faster R-CNN is a two-stage network,which firstly uses RPN network to generate several high-quality suggested regions,and then sends the suggested regions to Fast RCNN network for further target identification and positioning.The experiment compared the performance of Faster R-CNN training under different feature extraction networks.It was found that when the feature extraction network was Res Net101,the performance of the network was the best.Then,the effect of Anchor in different size range on network performance was compared and tested.The test shows that when the size range of Anchor is [48*48 96*96 192*192],the network obtains the best training result.Finally,this study compares the effects of different image enhancement methods on network performance,and the experiment shows that the data set expanded by image clipping and image cutting is the best way to improve network performance.Through the above three experiments,an appropriate feature extraction network was selected for Faster R-CNN,network parameters were adjusted and network training was optimized,so as to improve the performance of the network.The final trained network AP value was 0.658.(2)The brown planthopper detection algorithm based on R-FCN algorithm was studied.R-FCN algorithm was improved based on Faster R-CNN,which retained the structure of RPN network,added the position fraction graph in the core network,and replaced the full connection layer with the convolution layer,thus improving the positioning accuracy of the network and accelerating the speed of network operation.The experiment also selected the appropriate feature extraction network and Anchor size range for the R-FCN network,and improved the performance of the network through image enhancement.The R-FCN network also provides OHEM training mode.The experiment compares the difference between general training mode and OHEM training mode.After the optimization training,the AP value obtained by R-FCN is 0.762.Finally,the specific detection results of the optimal training model of R-FCN and Faster R-CNN were compared,and the algorithm with better comprehensive performance was selected for the detection and counting of brown planthopper.(3)A two-layer brown planthopper detection and counting algorithm was proposed.The first layer of the algorithm was used to detect the rice region in the image.The results of regional growth method,dajin method,k-means segmentation algorithm,fuzzy c-means segmentation algorithm and Faster R-CNN algorithm were compared in the experiment,and finally Faster R-CNN +VGGNet16 was selected to extract the rice region.In the second layer of the algorithm,the rice region extracted from the first layer was divided into two regions,then the trained R-FCN network was used to detect and count brown planthoppers.The experiment shows that the two-layer brown planthopper detection and counting algorithm can effectively improve the recall rate of the algorithm,and the recall rate decreases with the increase of the number of planthoppers,which is lower than that of the single-layer detection algorithm.Tests showrd that the two-layer brown planthopper detection and counting algorithm can effectively improves the algorithm’s recall rate,and the algorithm’s recall rate reached 60.44%.(4)The brown planthopper detection and counting system based on web page and server is developed.Users can upload images for detection and counting of brown planthopper,and the results of previous surveys can be searched on the "my detection" page,as well as the specific detection results of each survey. |
PDF Full Text Request