Delicate image segmentation based on cosine kernel graph cut

Abstract

The kernel graph cut approach is effective but highly dependent on the choice of kernel used to map data into a new feature space. This study introduces an enhanced kernel-based graph cut method specifically designed for segmenting complex images. The proposed method extends the RBF kernel by incorporating a unique mapping function that includes two components from the MacLaurin cosine kernel series, known for its ability to decorrelate regions and compress energy. This enhanced feature space enables the objective function to include a data fidelity term, which preserves the standard deviation of each region’s data in the segmented image, along with a regularization term that maintains smooth boundaries. The proposed method retains the computational efficiency typical of graph-based techniques while enhancing segmentation accuracy for intricate images. Experimental evaluations on widely-used datasets with complex shapes and fine boundaries demonstrate the effectiveness of this kernel-based approach compared to existing methods.