Superpixel Segmentation Based on Delaunay Triangulation

Superpixel segmentation is popularly used as a preprocessing in image segmentation and object recognition. However, none of effective superpixel segmentation algorithms have been proposed to highlight the contour features of the salient object in image, and they are difficult to describe the irregular superpixel boundaries. That will make a bad impact on subsequent image processing. In this paper, we present a novel algorithm to produce superpixel based on Delaunay Triangulation, and the Difference of Gaussian (DoG) feature points are used as nodes to build initial superpixels. This method will reduce the redundant superpixels of the image's smooth regions. To further avoid redundant segmentation, firstly, we obtain the gradient image by Roberts operator for texture features, and the RGB color space is converted into the YUV color space for color features. Secondly, we calculate the local contour orientation of the neighborhood region of superpixel boundaries. Thirdly, we compare the superpixel boundaries orientation with the local contour: if the orientations are consistent, we retain the boundary; if not, we delete them. Likewise, if the average colors on both sides of the boundary are different, it will be retained. Experimental results on the Berkeley Segmentation Dataset (BSD) show that the superpixel segmentation algorithm in this paper can effectively suppress the redundant superpixels and the superpixel boundaries can draw the contour of the salient object clearly. Moreover, our method can self-adaptively adjust the number of superpixel based on image texture and HSI color features.