High Resolution Remote Sensing Image Classification Based On Bag Of Visual Word

In recent years, the rapid development of remote sensing technology, the characteristics of remote sensing images tend to be large amount of data, high dimension, high resolution, it contains the feature information is becoming more and more rich, so how to efficiently in a large amount of data to extract the needed information and classified, it become the research hotspot in the field of remote sensing images. Compared with the traditional remote sensing image classification method based on pixels, object-oriented bag-of-visual-words model overcame the features to describe a single, the algorithm effectively solves the "semantic gap"between the high and low level features and other issues. Therefore has been widely research and application.Classic bag-of-visual-words model in key technologies and difficulties in high resolution remote sensing image classification, this paper carried out in-depth study.Put forward a complete set of effective remote sensing image classification scheme.The first image is segmented into homogeneous object, extract the object's characteristics to generate visual words; then, the screening of visual words, choose words and correlation between big, redundancy between words and words of small subset of the dictionary; the final choice SVM classifier achieve image classification.In this paper, the main content and innovation points are as follows:(1) In view of the traditional image segmentation algorithms are susceptible to noise interference and over-segmentation phenomenon is serious problem, this paper proposes a combination mark watershed and region growing image segmentation method. The experimental results show that: the algorithm can not only filter out noise but also better retain boundary information, and can restrain the over-segmentation phenomena.(2) Aiming at high and low layer characteristics exist in the "semantic gap"problem, the bag-of-visual-words model is introduced into the remote sensing image classification, extraction and fusion of many characteristics, enhance the feature describing ability of the algorithm, breaks the limitation of the traditional classification algorithm.(3) Because there are a large number of redundant information in the visual dictionary, lead to the increase of computational complexity and decline in classification accuracy, so to choose words. MRMR algorithm considering the correlation of the word and redundancy, but cannot reflect different words on the classification of the contribution. The word weighting parameter is given by the ReliefF algorithm to establish the dictionary discriminant function, this paper puts forward the improved mRMR criterion of visual word selection algorithm.(4) Design the SVM classifier for high resolution remote sensing image classification experiment, the key parameters involved in the algorithm are experimentally optimized. The validity and advancement of the algorithm are verified by experiments. Compared with the pixel-based classification algorithm, the overall classification accuracy of this algorithm is improved by 33.3%, Kappa coefficient is increased by 0.37.