A Hybrid Colour Image Segmentation Scheme


An autonomous region- and pixel-based segmentation scheme has been developed that efficiently and robustly partitions a colour image into different regions that are homogeneous with respect to colour. As with human visual perception, image segmentation is an important aspect of any type of image or scene analysis. In the one case, humans use their visual sense to effortlessly partition their surrounding environment into different objects to help recognize the objects, guide their movements, and for almost every other task in their lives. In the other case, image segmentation is usually the first task of any artificial image analysis process and all subsequent tasks, such as feature extraction and object recognition rely heavily on the quality of the segmentation.

The colour image segmentation scheme is based on the HSI colour space. The benefits of considering the hue, saturation, and intensity colour values of pixels and regions include good compatibility with human intuition, separability of chromatic values for achromatic values, and the possibility of using only one colour feature (hue) for segmentation. All three of these benefits are taken advantage of in the segmentation scheme.

The hybrid-based colour image segmentation scheme combines both region- and pixel-based techniques. A pixel classification algorithm is used to categorise the pixels in the image as either chromatic or achromatic. This algorithm alleviates the problem encountered when an achromatic and a chromatic pixel need to be compared in terms of their colour value. A local variance seed determination algorithm is then used to find seed pixels in the image. Seed pixels are used in the region growing algorithm as starting points. Finding the 'best' seed pixels encompasses finding the pixels that are dominant in colour and predominantly in the spatial centre of the region.

The region growing algorithm starts with the set of seed pixels and from these grows regions by appending to each seed pixel those neighbouring pixels that satisfy a homogeneity criterion.  The Cylindrical distance metric is used in the homogeneity criterion to compare the colour similarity between pixels. The HSI colour space is represented in a cylindrical coordinate-type space and, therefore, correlates well with the Cylindrical distance measure. The region growing algorithm is a region-based technique.

After regions are grown they are further processed with a region merging algorithm. Regions are merged if they are: (1) joined and similar in colour; or (2) similar in colour but spatially not connected. Colour distance measures are used to test the colour similarity between regions. Once again, the Cylindrical distance measure gives the better results.