Abstract:

In order to avoid the high computational-complexity inherited in traditional texture extraction method, a novel, simple, yet effective textural feature extraction method for high resolution remote sensing image classification is proposed in this paper. First, the original texture extracted from local image patches are projected into the compressed sub-space using the random projection technique. Then, the texture dictionary which represents local features is learned with k-means in the compressed domain for each class. Finally, the visual word map is formed by coding every texton in the samples to the nearest word in the texture dictionary, and then the histogram of the visual words map and the second moment of the words are fused as the final textural feature. The propose method is proved to be effective for texture representation and improving accuracy for high remote sensing image classification by two groups of experiments.

Key words: High resolution remote sensing, Random projection, Texture feature, Texton, Visual words, Classification