Studying the Impact of the Color Representation Choice on Segmentation Results by the Seeded Region Growing Algorithm

Abstract

With the adaptation of traditional image processing methods to answer to the requirements of analyzing more complex representations than the standard gray scale portrayal, a growing interest is directed towards color image processing. Many color spaces are suggested in the literature, each serving a defined purpose and generating an output image with distinguishable perceptual differences. One other major difference between existing color spaces is the meaning each of them gives to the color component that makes certain operations and transformations possible on a representation and not on another. This raises the question whether certain processing may score distinct results on different color spaces. In this paper, we investigate the impact of RGB and HSV color representations on one the most famous segmentation methods: the Seeded Region Growing algorithm (SRG). The implemented method follows a three step process: 1) The automated seed selection, based on two criteria, 2) The region growing phase, that will gather neighboring pixels and seeds into regions according to a similarity rule guided by the Euclidean distance, and at last, 3) the region merging phase will proceed to overcome the over segmentation issue and enhance the meaning of the defined regions will improving the segmentation accuracy. Three metrics from the literature were used to assess the performances of our algorithm on both color spaces conforming to similarity, distortion and noise. The segmentation results were compared by combining and averaging the performance measures from an images sample from the Berkeley dataset. The algorithm showcased more accurate results and consumed less execution time in the HSV color space compared to the RGB one. Post processing steps were tested and included in the RGB model to enhance the segmentation outcome.