A graph based approach to hierarchical image over-segmentation

• A new approach to segment partitioning for superpixel segmentation is proposed.
• A new objective function optimized with graph cuts is proposed.
• The algorithm produces compact and regular superpixels of high quality.
• Comparable results with contemporary algorithms are achieved.

The problem of image segmentation is formulated in terms of recursive partitioning of segments into subsegments by optimizing the proposed objective function via graph cuts. Our approach uses a special normalization of the objective function, which enables the production of a hierarchy of regular superpixels that adhere to image boundaries. To enforce compactness and visual homogeneity of segments a regularization strategy is proposed. Experiments on the Berkeley dataset show that the proposed algorithm is comparable in its performance to the state-of-the-art superpixel methods.