A Robust Color Watershed Transformation and Image Segmentation Defined on RGB Spherical Coordinates

Introduction

Image pre-processing and image segmentation are key steps on robotic vision. On the one hand, humanoid robots are in continuous movement, changing of sceneries, changing of point of view, and in all cases illumination conditions could be unstable and different, e.g. a robot can go by walk from a corridor with natural illumination to a room with tungsten illumination. In this case, a robust image pre-processing is key in order to normalize the image respect to the illumination. On the other hand, when the image is already normalized, the following step, is the information extraction. The most used is the segmentation process. This step divides the image in some regions such that these regions can be identified as objects in the scene. For this work, we have found robustness respect to the illumination through spherical coordinates, and for segmentation we will use a watershed transform with a region merging directed by the image chromaticity.

Color images have additional information over gray scale images that may allow the development of robust segmentation processes. There have been works using alternative color spaces with better separation of the chromatic components like HSI, HSL, HSV, Lab (Angulo & Serra, 2007; Hanbury & Serra, 2001) to obtain perceptually correct image segmentation. However, chromaticity's illumination can blur and distort color patterns. Color constancy is the perceptual mechanism which provides humans with color vision which is relatively independent of the spectral content of the illumination of a scene. It is the ability of a vision system to diminish or, in the ideal case, remove the effect of the illumination, and therefore “see” the true physical scene as the invariant to illumination changes. To obtain color constancy, one approach consists in the estimation of the illumination source chromaticity followed by the chromatic normalization of the image. There are several approaches in the literature to achieve this goal (Tan, 2003; Yoon, 2005; Toro, 2008) assuming a uniform chromaticity of the illumination all over the scene. Other approaches try to obtain segmentation procedures which are inherently robust to illumination effects (Mallick, 2005; Zickler, 2006). The segmentation method proposed in this paper has inherent color constancy due to the color representation chosen and the definition of the chromatic distance.

Color constancy is closely related to the response of the gradient operators (Geusebroek, 2003). Regions of constant color must have low gradient response, while color edges must have a strong gradient response. Image segmentation methods based on spatial gradients need a correct definition of the spatial color gradient and unambiguous contour definition. In fact, formulation of watershed segmentation methods in color images is still an open research issue (Aptoula, 2007). A straightforward but inexact approach is the independent application of the watershed segmentation on image channel (Tarabalka, 2010). This approach loses chromatic information, and has difficulties merging the subsequent independent segmentations into one.