Unsupervised texture segmentation applied to natural images containing man-made objects

Abstract

This paper presents a region-based unsupervised segmentation for natural images containing man-made objects. We propose a texture feature extraction to obtain more discriminating features. Statistical Geometrical Features (SGF) are used as texture features. The SGF of the original image and the smoothed image obtained from an anisotropic edge-preserving diffusion are combined for segmentation use. We also propose a modified segmentation algorithm which performs segmentation in four stages: hierarchical splitting, local agglomerative merging, global agglomerative merging and pixelwise classification. Local agglomerative merging combines segments locally, which will greatly reduce the time cost. We make some experiments to demonstrate the effectiveness of the proposed technique in the segmentation of natural images containing man-made objects. The reduction of computation time is also provided.