Study Method On Extraction Building Of High Resolution Image Based On The Bag Of Words Model

Geographic national condition belongs to a part of national condition that makes relevance and analysis of all kinds of conditions by way of spatial properties, including natural environment and resources, science and technology education, economic development as well as international environment in order to reveal the spatial and temporal evolution of social development and internal relationship of comprehensive national conditions. However, the basic information is supported by the data of remote sensing image. The successful launch and operation of the two satellites—GF-1 and GF-2, provide sufficient data research for the census of our country’s national conditions. With the help of these two satellites, the high-resolution photography can be fast obtained and the period of refreshment of image data is much shorter compared with before. However, the traditional method of artificial visual interpretation cannot cater to the needs of refreshment of modern information. Therefore, how to access a large number of high resolution image data fast and efficiently has become a critical problem that needs to be solved urgently. This thesis attempts to aim at solving automatic and efficient extraction of building information combined with calibration method based on BOVW and method for extracting edge based on Grab Cut in order to finish the automatic extraction of high resolution image in the building information and has achieved some experimental results.This thesis introduces two calibration methods based on BOVW. One is a building method based on spatial pyramid bag of words. This method takes advantage of SIFT as a characterized word by way of K-means to get a dictionary combined with the theory of spatial pyramid to build imaging features of model of three layers of word frequency histogram features and use SVM to make a classification. The results prove that the method has achieved better effects by using high resolution images respectively with 2 meters and 0.5 meters to make experiments. The other is a building method based on the MRF characterization of BOVW. This method is combined with the thoughts of MRF theory and combines the neighboring pixel with several pairs of visual words according to a certain order, including the co-text information and the lessen of primitive information. The results illustrate that the whole time in the second method has sped up nearly twice despite of the lower accuracy compared with the first method.The method for extracting edge based on Grab Cut is on the basis of the building information image to obtain the refined boundary information of buildings. This method uses the calibrated images which are connected with calibrated building information from the original one to intercept the small images which are cut by Grab Cut method in order to obtain the whole boundary information in the image building. The results prove that the method has achieved good results by using high resolution images respectively with 2 meters and 0.5 meters to make experiments.