Robust watershed segmentation using the wavelet transform

The watershed transform has been used for image segmentation relying mostly on image gradients. However, background noise tends to produce spurious gradients, that cause over-segmentation and degrade the output of the watershed transform. Also, low-contrast edges produce gradients with small magnitudes, which may cause different regions to be erroneously merged. In this paper, a new technique is presented to improve the robustness of watersheds segmentation, by reducing the undesirable over-segmentation. A redundant wavelet transform is used to denoise the image and enhance the edges in multiple resolutions, and the image gradient is estimated with the wavelet transform. The watershed transform is then applied to the obtained gradient image, and segmented regions that do not satisfy specific criteria are removed.