Hierarchical Image Segmentation Based on Multi-feature Fusion and Graph Cut Optimization.

The task of hierarchical image segmentation attempts to parse images from coarse to fine and provides a structural configuration by the output of a treelike structure. To deal with the challenges of keeping semantic consistency in each level caused by the variable scale of different objects in image, this paper proposes a hierarchical image segmentation approach guided by multi-feature fusion and energy optimization. We transform the image into a region adjacency graph (RAG) by superpixels and design a bottom-up progressive merging framework based on graph cut for a hierarchical region tree. A multiscale structural edge is designed as a feature map for mapping to the hierarchical levels, while we conduct salient map and object window as a weakly-supervised prior during the optimization process. Experimental results demonstrate that our approach gets a better performance in semantic consistency while has an encouraging performance compared with some state-of-the-arts.