To overcome the shortcomings such as the existence of over-segmentation in and the great influence of noise on the traditional watershed segmentation, an improved novel watershed image segmentation method is proposed. Firstly, the open-close reconstruction filter is adopted to remove the image noise. Secondly, multi-scale structure elements are used to calculate morphological gradient which can overcome the shortcomings of other conventional gradient operators and produce a gradient map suitable for watershed segmentation. Furthermore, the morphological gradient is modified by viscous morphological operators to remove the most irregular local minimums. After the standard watershed transforming, the region merging method based on neighbor regions edge value is employed to improve the segmentation result. Experiments show that this method can not only effectively remove the noise but also avoid the over-segmentation of watershed and preserve the positions of regional contours.