Feature extraction for man-made objects segmentation in aerial images

Abstract

This paper describes a new aerial image segmentation algorithm by level set. The non-subsampled contourlet transform (NSCT) is selected to represent the image features for the classification of man-made object. NSCT can avoid pseudo-Gibbs phenomena around singularities during the process of image denoising, owing to the properties of shift-invariant. NSCT also enriches the set of basis functions which makes it possible to extract some critical signal features. The optimization of basis selection is applied in the NSCT to ensure the decomposition based on the maximum information content. Another kind of feature named RMSerror is also extracted from fBm model. Then a modified Mumford-Shah model which comprises the NSCT features and RMSerror features extracted from the aerial images is built to segment the aerial image by a necessary level set evolution. At last the proposed method is proven to be effective by the results of experiments.