Research On Statistical Methods Of Spruce Quantity Based On Visible Light Nusery Image Of Unmanned Aerial Vehicle

As an important part of nusery production,inventory statistics provides decision data for nusery production plan.At present,the inventory statistics of nusery mainly rely on manpower,which has the problems of low efficiency and unstable precision.It is urgent to develop an automatic accurate inventory statistics system.This paper takes spruce as the research object,and discusses how to implement the counting method of visible light image in spruce nursery from the four aspects of acquisition,splicing,segmentation and counting of spruce image.The main work is as follows:1)470 aerial images of spruce are selected and labeled,and the stitching of spruce images is obtained by SIFT algorithm.This paper analyzes the relationship between the flight height of UAV and the definition of image,and then draws up the flight plan of UAV.A large number of nursery images were collected.470 semantic segmentation annotation and 34484 slice count annotation were made.The dataset was expanded 4 times by the data enhancement algorithm.To compensate the limited vision of UAV camera,spruce images is stitched by extracting the SIFT features of the spruce image,matching feature points and fusing the images.2)Three spruce image segmentation algorithms based on kmeans,SVM and FCN are designed.The RGB,HSV color features and three LBP texture features of the nursery image are extracted,and the nursery image segmentation algorithms based on Kmeans and SVM are implemented respectively.By analyzing and comparing the effects of various feature combinations on the algorithm,the best segmentation model is selected.In order to obtain a higher-precision segmentation model,three FCN models are designed according to the concept of FCN.Their feature extraction layers are built on Alex Net,VGG16 and VGG19,respectively.Upsampling and feature fusion are also constructed to obtain complete model.Models are tested and analyzed in the test set.In the end,the VGG19-FCN model achieved a 0.75 MIo U in spruce semantic segmentation test,which far exceeding the segmentation models based on SVM and Kmeans.3)Three kinds of spruce counting algorithms based on Hough circle detection,slice counting CNN and Mask RCNN are proposed.To reduce the difficulty of processing,median filter,opening and image filling are used to enhance the segmented image.In this paper,we first use Canny edge detection and Hough transform to implement the spruce counting algorithm based on Hough circle detection,then use the connected region extraction and CNN to implement three kinds of spruce counting algorithms based on slice counting CNN.The feature extraction layer of slice counting CNN is constructed by Alex Net,VGG16 and VGG19 respectively.The softmax classifier in the CNN is removed to adapt to the counting task.The results show that the Hough circle counting algorithm has higher recognition accuracy for spruce images with a small number of plants,but it is less effective in processing pictures with a larger number of spruce.The counting convolutional neural network constructed with VGG19 can achieve higher precision spruce number counting based on the segmentation results of the VGG19-FCN model.In order to achieve a higher-precision end-to-end counting model,a Mask RCNN instance segmentation network with Res Net101-FPN as the feature extraction layer was constructed,and the spruce dataset was used for training.The model is completed by constructing modules of featrue extraction,RPN,bounding box regresion,classification and segementation.When counting,the counting result is obtained by counting the number of spruce instances in the test picture.Experiments show that the algorithm has the highest counting accuracy,which is 95%.According to the above research,the VGG19-FCN segmentation model has a high accuracy in spruce image segmentation,and has strong robustness for outdoor interference;the spruce counting algorithm based on Mask RCNN has a high counting accuracy in the counting task,and can also effectively identify the high-density spruce image.The algorthmes provides a solution to the problem of seedling statistics.