Research On Residents Extraction Of RS Images Based On Texture Features

With the development of sensor technology, airborne and spaceborne platform technologyand data communication technology, remote sensing technology has entered a new stage. And itis capable of accessing various types of earth observation data dynamically, rapidly, accuratelyand by multiple means. Extracting spatial information from remote sensing imagery has becomean important approach for geographic information acquisition. And the exacted information iswidely used in the national economy development and military surveying and mapping support.In the thesis, the focus is put on the studying of image texture based residents extractionalgorithm, the core of which is resident direction detection, the subsequent feature extraction andother related techniques.The main contents and innovations in the thesis are listed as follows:1. Introduction was given on the concept, significance and difficulties of remote sensingimage understanding, and summarization was made on the basic methods, research status anddeveloping trend of ground object extraction. Analysis was carried out on the emphases anddifficulties of residents extraction. And the research scope and basic ideas in the thesis wasdefined.2. Analysis was made on the expression forms of the residents, vegetation, water and theother typical features in remote sensing images, and qualitative description was given on theirtexture characteristics in the image. The expression methods, which are based on differenttexture indexes, and distance measurement algorithms were summarized, and an extractionprocess based on texture feature was then designed.3. A detailed analysis was given on the principle and method of spectrum analysis, Gabortransform, co-occurrence matrix, multi-scale autoconvolution and Tamura texture. Based on themain direction of residential texture, feature construction strategies and extraction schemes weredesigned.4. One residents extraction method based on spectrum analysis and one Gabor transformmethod based on residents texture direction were brought forward. Based upon the contributionof the above three features in the image space, different weights were assigned to the features toconstitute a feature vector in the first method. The new vector is characteristic of low correlationand strong differentiation among different features. The feature vector was reduced from40dimensions to2dimensions, which solved the bottleneck existed in huge filter groupcomputation, and at the same time, enhanced the ability of feature focusing analysis.in he secondmethod 5. One gray co-occurrence matrix construction method based on residents texture directionwas put forward and the construction method of multi-scale co-occurrence matrix was improved.The co-occurrence matrix direction was determined by two main texture directions, which is thestep ratio of pixel couples in the vertical and horizontal direction. Meanwhile, according to thedirection of pixel couples, the sliding window is changed from the conventional square torectangular and pertinence and validity of feature description is greatly improved. In the secondmethod, the first main texture direction was rotated to horizontal direction and then sampled bywavelet transform. On various scale low frequency,0oand90odirection high frequency,corresponding co-occurrence matrix was constructed. Combining the physical meaning andcorrelation analysis of features, measures of low redundancy and low correlation were selectedand then used for residents extraction.6. Based on Tamura contrast and MSA histogram, a residents extraction method wasproposed. MSA histogram features were compared, analyzed and then used for residentsextraction form remote sensing images. Since MSA histogram features were sensitive to noise,the measure of Tamura contrast was introduced to improve the anti-noise performance andprecision of the algorithm.7. According to the improvements and innovations of our algorithms, experiments weredesigned and analysis was made. Comparing with the results obtained by the original and theother algorithms, the effectiveness of our algorithms was verified.