Study on Image-segmented Classification

摘要

There are many studies about classification algorithm and the development of understanding systems. Most of these studies for image classification are on condition that the same kind of objects has the similar spectral features and the different kind of objects has the different spectral features. Generally, the images are very complicated because of sensors. Sometimes, the spectral features are same for different objects, while the spectrum features are different for the same objects. Obviously, this hypothesis will effect the classification accuracy. How to solve this problem? This paper presented a image-segmented classification method. In this method, the image was firstly divided into several segments or areas according to the spectral features, then the different training schemes was used for classification of the different segments; finally, all the results can be combined automatically into a file. The experiment for this method is land use classification for a image coming from two different scenes. The results are 6 classes which classification precision are more than 80% and 3 classes more than 90%. However, the classification for the whole image at a time, only two classes which classification precision are more than 80%. The experiment proves that image-segmented method can improve the quality of image interpretation either in accuracy.