A DISCRETE LOG DENSITY EXPANSION BASED APPROACH TO IKONOS IMAGE CLASSIFICATION
Donghui Yan, Peter J. Bickel, Peng Gong · 2006
We approach the problem of Ikonos image classification through a type of image statistics - GLCM. By viewing the GLCM as a random matrix, we make connections between texture and an expansion of the logarithm of the joint density of entries in the random matrix. We propose a bottom-up approach to characterize texture for the purpose of classification by selective inclusion of higher order terms in the log density expansion. This approach is implemented with a greedy forward stagewise procedure and starts with a model that includes all first order terms. We obtain estimate for the coefficient of terms in the log density expansion under conditions that are justifiable empirically. Experiments on a real data set - a set of Ikonos images for 9 different texture classes, show that our approach outperforms a procedure that uses domain expert knowledge, and achieves results close to that obtained by AdaBoost.M1 on a naive Bayes classifier built from the base model our greedy procedure starts with.