Color Image Segmentation Using Improved Graph Cuts

A novel unsupervised color image segmentation method using graph cuts with multiple components is proposed，which can overcome the problem of the higher computational complexity caused by more labels during inferring by graph cuts. First，the quaternion cut-off window feature and CIE Lab color feature of a given image are extracted and fused based on the gradient information of the image. Then the segmentation is formulated as a labeling problem and solved by an iterative process based on graph cuts and maximum likelihood(ML) estimation. At each iteration，the connected regions in a segment are handled as sub-components of the segment instead of relabeling them with unique labels. In doing so，the number of labels does not increase，and thus the computational complexity can be reduced during inference by graph cuts. Finally，the segmentation result is obtained after removing some weak edges. Experimental results and theoretical proof demonstrate the good performance of the proposed method.