Multiscale image segmentation using wavelets and watersheds

We propose a new multiscale segmentation technique based on wavelet decompositions and watersheds. The wavelet transform is applied to the image, producing detail and approximation coefficients. The watershed is applied to the approximation image at a certain resolution 2^J, and projected up to higher resolutions using the inverse wavelet transform. At lower resolutions, the segmentation captures large relevant objects, while at higher resolution more details are obtained. Since downsizing is inherent to the wavelet transform, watershed segmentation is applied to a smaller image, demanding less computational time. The proposed technique appears to be robust for noisy images, even when the amount of noise is very large.