Reserch Of High-resolution SAR Images Feature Extraction And Pattern Retrieval

In the medium-low resolution SAR images, the classification methods and retrieval methods which are based on pixels can obtain good results. However, with the the spatial resolution increasing, the bigger the amout of data of high resolution SAR image, the more obvious the object feature and the information is richer. Meanwhile, pixel-based image analysis does not consider the relationship of spatial organization between pixels. Then the traditional pixel-based classification method will make the results ambiguity and uncertainty.We make the image blocks containing one or more contents be local modes. This paper proposes a method for image interpretation based on local mode, which is used to replace the method based on pixels. Pattern retriavel is, according to the local modes, to retrieve a local mode which has the nearest content with it in a large image, which makes find specific local image quickly and accurately under the rich imformation of image circustance. The thesis focuses on researching the feature extraction method and the classification model of local modes. And the thesis gives two applications of local modes classififcation which are ground objects types extraction and local modes retrieval. Therefore, this thesis will make use of pattern retrieval method to solve the high-resolution pattern classification.The work of the paper is as following:1.Based on the current feature extraction and pattern retrieval theory of SAR image, combining with the character of high resolution SAR image, analysis the exsting problems of current feature extraction. Extract the local modes of high-resolution SAR images, and calculate the SIFT features of each local mode. Then make use of the bag of words to count its integrated characteristic, combined spatial pyramid model and GLCM theory which shows SIFT spatial distribution of feature points. Finally, make use of the SVM to classify the local modes, which belong to farmland, rivers and urban categories.2.Based on the classification of local patterns, research the extracting method of the ground objects types based on the local patterns classification. Only process the interesting local modes to extract the ground object type. And map them back to the original image to extract the whole ground object type. Then make use of the river and city to verify the model.We ultilize the morphological method and soble operator to detect the river in the local patterns wich contain river. And map them back to the original image according to the partition rule of the local patterns. As for the city extraction, we divide the local modes which belong to the city into mirco modes(smaller local mode), using the morphological profile and SVM to classify the micro modes into urban and non-urban. Make the micro modes which belong to the urban mapping back to the original image to obtain a crude city results. Then two schemes are given out to present fine urban result. One is based on the Gaussian blur edge to remove the areas which don’t belong to the city area. The other is based on likelihood function of SIFT feature points. The SIFT points belonging to urban are more than them in other area s. Then remove the area which don’t belong to urban to obtain accurate urban extraction results. Experiments results show that the first option has higher efficiency, simpler and more quickly than the second option, but the result is relatively rough. The efficient of second scheme is lower while its accuracy is higher.3.On the basis of local pattern classification, propose a local pattern retrieval approaches of Gabor feature and TF-IDF weighting technique. Only processing the local patterns which belong to the same class as the same as the query local pattern, the average retrieval time without local pattern classification process is 0.75 s.