A Spatial Constraint Image Segmentation Algorithm Based on Block Clustering

Abstract

Image segmentation is a key preprocessing step for object recognition and has a profound effect on the subsequent classification and recognition. Visual spatial clustering based segmentation is a commonly used method in image segmentation, which clusters pixels using visual descriptors by space similarity measure. It can achieve good results in simple image segmentation with less noise. This paper presents a segmentation method based on spatial position constraint of the pixel. The image is divided into overlapping rectangular blocks. By iteratively clustering these blocks with typical visual features and splitting the blocks with worse visual consistence, the spatial constraint information is added to the clustering process implicatively. The method is still unsupervised learning algorithm essentially by no guidance information provided beforehand. Experiments using real images are presented to show the efficiency of the proposed algorithm with better segmentation results than K-means.