Image Classification of Marine Landmarks Based on Evidence Theory

Island classification is a key element of scene matching navigation in the sea area. However, the image classification methods face the problems of inconsistent distribution structure, uneven size and lack of stable features of islands. In order to solve these problems, we define three types of island, the isolated island, the large island and the multi-island. This paper considers the pyramid decomposition to perform multi-scale analysis, and uses the histogram of oriented gradient and the local binary pattern algorithm to extract the stable features of the island images at different scales, then these feature vectors of each scales are classified by support vector machines. Furthermore, the evidence theory is introduced to fuse the classification results of single classifier on each image scale. The island database is obtained by Google Earth satellite images, which covers all islands in South China Sea and some of islands in Pacific/Indian Ocean. The experimental results on the satellite image database show that the classification accuracy of proposed method is 91.83%, and it is about 2% higher than single classifier methods.