Selective color processing via automatic color segmentation

Selective color processing is a color image processing scheme which is performed only on a particular color component in a color image. Extraction or isolation of the color component can be achieved via color image segmentation. In this dissertation, an automatic color image segmentation algorithm based on the modifications of the multi-scale clustering (MSC) algorithm is introduced for the purpose of performing selective color processing. The developed color image segmentation algorithm isolates homogenous color regions in a color image in an automatic manner without requiring the number of colors or color regions to be specified by the user. Although the original MSC algorithm can be used to generate the prominent prototypes in the color domain, it has limitations when it comes to using it for color segmentation. This dissertation presents modifications to the original MSC algorithm by using the Riemersma's color difference measure, variable step sizes, and a color space sectoring procedure. The modified algorithm is referred to as MMSC. The computational complexity is also reduced by introducing a restricted potential function computation approach. In addition, a multi-layered chrominance segmentation step in the color domain and a fine color segments merging step in the spatial domain are introduced to make use of the spatial information. The developed color segmentation algorithm has been tested using many real and synthetic images. Five color difference measures and a newly introduced color edge likelihood measure have been deployed to evaluate the goodness of color representation and segmentation. The color difference measures indicate that the MMSC color segmentation algorithm produces on average a lower color distortion as compared to the widely used c-means type of clustering algorithms. Furthermore, the introduced color edge likelihood measure indicates a closer match to color edges as compared to the c-means type of clustering algorithms. The evaluation performed on synthetic images demonstrates that the developed MMSC color segmentation algorithm is more tolerant to noise. Several applications are presented to show some of the uses of the introduced algorithm.