Color based image segmentation as edge preserving filtering and grouping

In this paper we contend that color based image segmentation can be performed in two stages: an edge preserving filtering stage followed by pixel grouping. Furthermore, for the first step, we introduce a framework under which many current filtering approaches (and some novel ones) can be classified. We present experiments where a new method, called Color Mean Shift, outperforms all the other methods in producing more uniform regions and preserving the edge boundaries. Then, applying standard clustering methods for the grouping step, we present extended experimental results for the Berkeley segmentation dataset of the combined filtering+grouping segmentation methods. We show that this new approach to image segmentation produces better segmentations compared to current state-of-the-art methods based on grouping only.