Design And Implementation Of Vegetation Recognition System Based On Aerial Mosaic Image

In recent years,the rapid development of UAV technology has promoted the rapid progress of low-altitude remote sensing technology.The advantages of UAV,such as convenient operation,high efficiency and low cost,make it widely used in forestry monitoring,disaster emergency response and other fields.Due to the small view angle,large amount of data and high resolution of UAV low-altitude aerial photography images,in order to get a complete view of the shooting scene,it is necessary to mosaic a large number of small view images into large scene images by mosaic technology.Dynamic monitoring of forest growth also requires classification and extraction of vegetation.Therefore,it is an urgent problem to select which algorithm to achieve accurate and efficient image mosaic and which method to use to achieve vegetation recognition.In the part of image mosaic technology,this paper systematically studies and summarizes the basic theory of image mosaic technology such as aerial image preprocessing,projection transformation and image fusion.Five widely used Kaze,Sift,Surf,Orb and Akaze algorithms are used for feature extraction experiments.Four indexes,stability,speed,repetition rate and registration accuracy,are used to evaluate the performance of the algorithm.Finally,SIFT algorithm with strong robustness,high accuracy but slow speed is selected.In order to meet the needs of practical application,some improvements are needed to speed up the SIFT algorithm.This paper adopts the following strategies to improve the speed of SIFT algorithm: first,to reduce the size of the image by performing descending sampling operation before extracting feature points from the image,and then to modify the detection range to 5x5 in the process of extremum detection of SIFT algorithm,so as to reduce the number of feature points and use grid-based motion.Estimation algorithm replaces the traditional RANSAC algorithm to screen out mismatch points.Experiments show that the improved SIFT algorithm improves the matching time of two images by about 1 s,and does not affect the final stitching effect.Vegetation recognition is essentially to extract and identify different types of vegetation regions after semantically segmenting the image.The application of deep learning in semantic segmentation can achieve automatic and accurate classification at the pixel level,so this paper chooses the deep learning framework to realize vegetation recognition.The process is as follows: labelme is used to label aerial images manually,and then the labeled images are expanded by rotation,adding noise and random clipping.70% of the data sets are taken as training sets.After constructing the model,the training set is input for training.The optimal model is obtained by repeated iteration training.The model can be predicted by using this model.Semantic segmentation of image is realized.Through vegetation identification,the purpose of monitoring the growth of forest land can be achieved and the management of forest farm can be facilitated.According to the actual application,in Microsoft Visual Studio 2010 environment,using C++ language programming to achieve the relevant algorithm modules,design and build the interface of vegetation recognition system based on aerial mosaic images,and integrate each functional module into the system to form a complete recognition system.The practical test shows that the system is stable in use,convenient in operation and can meet the project requirements.