Image segmentation based on merging of sub-optimal segmentations

In this paper a heuristic segmentation algorithm is presented based on the oversegmentation of an image. The method uses a set of different segmentations of the image produced previously by standard techniques. These segmentations are combined to create the oversegmented image. They can be performed using different techniques or even the same technique with different initial conditions. Based on this oversegmentation a new method of region merging is proposed. The merging process is guided using only information about the behavior of each pixel in the input segmentations. Therefore, the results are an adequate combination of the features of these segmentations, allowing to mask the negative particularities of individual segmentations. In this work, the quality of the proposal is analyzed with both artificial and real images using a evaluation function as case of study. The results show that our algorithm produces high quality global segmentations from a set of low quality segmentations with reduced execution times.