Research On UAV Image Building Extraction Technology Based On Image Object And Multi-feature Fusion

With the rapid development of drone technology,low-altitude aerial photogrammetry has developed rapidly in urban map updates and large-scale topographic mapping.The extraction of buildings is basic and important work of UVA image analysis.The traditional image classification methods that based on single pixel analysis of the image,such as supervised classification and unsupervised classification,sometimes cannot meet the accuracy requirements of classification when the classification objects are in a relative complex condition especially in the case of UVA images that is of much more details and complexity.Fortunately with the development of object-based concept,the classification accuracy has also been greatly improved.Using UVA images of a study area as test data,and adopting object-based image classification method and softs,the related digital image processing algorithms and the accuracy of integrating a variety of features in the classifier for image classification have been experimentally studied in this paper by a comparative study.The research approaches and results can be summarized as follows:1.Focused on the problems of high degree of data redundancy and many small features and blurry building boundaries in building extraction,the sobel differential operator is used to filter the images to highlight the image boundaries.2.A comparative study have done for the traditional pixel-based image segmentation technology and image object-based segmentation technology,and the advantages and disadvantages of the methods have been point out.Using ESP2 algorithm is used to obtain the optimal segmentation scale of the images.It is clear that the optimal segmentation scale parameter is 90,the shape parameter is 0.4,and the compactness is 0.5.3.The geometric parameter method is used to calculate the shape characteristics of the image.The adaptability and compactness of the matrix are verified through experiments to highlight the outline information of the building,in order to better extract the buildings.The eight gray level co-occurrence matrix method in e Cognitionto extract the texture features of the image.By comparative experimental study,it is clear that the information entropy in GLCM can better represent the texture of the features feature.4.The K nearest neighbor classifier(KNN)and support vector machine(SVM)approaches have been used to classify the UAV images.The Kappa coefficient of the KNN classification result is 0.88,and the SVM classification result is 0.87.When combined with the sobel operator,the classification accuracy of KNN is improved to0.96,and the classification accuracy of SVM is improved to 0.90,which shows that the sobel operator can effectively enhance image boundaries and improve classification accuracy.