A new color image segmentation algorithm based on watershed transformation

A new color segmentation method is presented in this paper. The method is specified for color images that have both large and small objects, and objects with both step and ramp edges. Scanned pages of color magazines and newspapers are the examples of this kind of images. Watershed transformation algorithm is the basis of the proposed method. Our method incorporates the original multi-scale analysis that allows to segment edges of different slope. This analysis uses fine-to-coarse strategy and prevents the already detected sharp edges from smoothing while moving to coarser scales. In the same time the introduced algorithm allows to detect ramp edges successfully at coarse scales. For fine scales we propose a special gradient operator and a modification of watershed transformation for small objects segmentation.

1. Introduction

Today color image segmentation is a rapidly developing area of digital image processing [1]. Color image segmentation attracts more attention due to the increasing computational potentialities of personal computers and possibility of their usage for color image processing. Nevertheless, this problem remains less investigated than segmentation of grayscale images. There are many application areas for color image segmentation, one of them is OCR (Optical Character Recognition) systems. Segmentation is the first stage in operation of OCR systems. OCR systems usually apply binarization methods for segmentation of grayscale images. However, application of these methods for color images is impossible in general. For this reason, development of segmentation methods for color images is a very topical problem.