Centroid Neural Network for Clustering of Line Segments

Abstract

An approach for an efficient clustering of 3D line segments based on an unsupervised competitive neural network is applied to a set of high resolution satellite image data in this paper. The unsupervised competitive neural network, called centroid neural network for clustering 3D line segments (CNN-3D), utilizes the characteristics of 3D line segments. Successful application of CNN-3D can lead accurate extraction of rectangular boundaries for building rooftops from an 3-D edge image which is considered as challenging and difficult because 3-D line segments are often contaminated with various noises obtained during stereo matching process. Experiments and results show that the proposed CNN-3D algorithm can group 3D line segments and the resulting 3D line groups can be successfully utilized for detecting rectangular boundaries for building detection.