Rapid, man-made object morphological segmentation for aerial images using a multi-scaled, geometric image analysis

Abstract

This paper describes a man-made object segmentation method for aerial images based on a modified watershed segmentation algorithm. Our segmentation procedure includes three steps: (1) a multi-scaled geometric image analysis of aerial images by the non-subsampled contourlet transform (NSCT) method, (2) watershed segmentation, and (3) region classification of man-made objects. First, background of multi-scaled geometric image analysis is introduced briefly, and NSCT is used to represent the features for the purpose of man-made object segmentation. Thanks to the properties of NSCT, it not only avoids pseudo-Gibbs phenomena around singularities in image de-noising with regard to shift invariance, but it also enriches the set of basis functions, which makes it possible to extract orientational contour of man-made objects more effectively. In the NSCT decomposition step, the best basis selection is employed for ensuring maximum information content. Second, the “texture gradient” of combined features is calculated based on the first NSCT decomposition step and the resulting best basis selection, afterward the watershed transform is applied. According to their feature values, the aerial images are divided into several homogenous regions. Third, the fractional Brownian motion (fBm) model is used to determine the man-made object regions. Last, the experimental results show that the outcome of man-made object segmentation becomes more continuous and satisfying as a result of the homogenous texture-regions extraction and the modified watershed procedure.