Tree Image Object Detection And Segmentation Method

The image information of standing trees directly affects the estimation accuracy of breast diameter,tree height,crown width,volume of forest resources,etc.It is very important for the informationization of forestry resources in China to obtain image information quickly and accurately.Aiming at the shortcomings of the existing tree detection and segmentation methods,this paper studied the object detection and segmentation method of tree images.The main work is as follows:(1)Standing tree image trunk detection framework: A model structure was proposed to rapidly detect trunks based on a convolutional neural network and transfer learning.I introduced the object detection knowledge of the yolov3 trained on the Pascal VOC2012 dataset by fine-tuning parameters.The classification and boundary prediction of the tree trunk object were realized by an improved multi-scale prediction strategy.(2)Trunks identification method: Five kinds of classification models such as Inception＿v3,vgg＿16,Res Net＿50,Res Net＿101,and Res Net＿152 are used for the trunks dataset to realize trunks identification.In this paper,the rapid detection method of trunk and the identification method of trunk are combined to accurately obtain the position and type of tree in the standing tree image.(3)Standing tree segmentation methods: This paper mainly adopts two kinds of segmentation methods: one is the interactive standing tree image segmentation based on Graph Cut algorithm,which is used in the object detection.The foreground and the background were marked internally,thereby obtaining each of the standing tree segmentation images in a single photo.The second is adaptive Mean-Shift algorithm for standing tree image segmentation.First,I abstracted the original tree image using bilateral filtering and image pyramid operations from multiple perspectives,which reduced the influence of image background information and tree canopy gaps on clustering.To calculate bandwidths,spatial location and gray scale features were introduced,obtained by step detection and the insertion rule method,respectively.Bandwidths determined the size of the Gaussian kernel function and were used in mean shift clustering.The flood fill method was then employed to fill the results of clustering to highlight the region of interest.(4)Standing tree counter extraction method: Traditional digital image processing techniques(such as grayscale transformation,edge detection,mathematical morphology processing,etc.)were used to obtain the standing tree edge information.The test results are as follows:(1)The results of tunk detection method showed that the average recall rate was up to 93.60%,the average misclassification rate was as low as 4.98%,and the average detection time was 0.30 s.(2)The results of the trunks classification method show that the recognition accuracy is as high as 94.67%.(3)In addition,the standing tree image segmentation experiment showed that the average segmentation rate of the standing segmentation method based on Graph Cut algorithm was 5.62%,the false positive rate was 4.49%,and the false negative rate was 4.33%,which was better than the OTSU,K-means and C-V segmentation method.The image segmentation results of the adaptive Mean-Shift algorithm based on image abstraction showed that the average segmentation accuracy,average over-segmentation rate and average under-segmentation rate of thecanopy were 91.21%,3.54%,and 9.85%,respectively.The corresponding average values of the trunks were 92.78%,8.16%,7.93%.It can be seen from the quantitative analysis that the standing tree detection and segmentation method has high precision and can be applied to forestry resource survey.