Image segmentation utilizing wavelet-based spatially adaptive kernels

Most multiresolution methods of segmenting images have hitherto resolved an image into successively coarser scales and discarded high-pass edge information. We present a robust and computationally tractable method for the automated segmentation of images by using spatially varying kernels derived from multiscale edge information of the image. We include examples of segmentations of synthetic and real images that demonstrate the performance of the algorithm in preserving fine detail and edge information in the segmentation maps while being robust to heavy noise.