Analyzing the Mallat Wavelet Transform to Delineate Contour and Textural Features

Analyzing natural scenes is made difficult when both contour and textural features are present. The problem of building suitable contour models from such images is compounded given texture region segmentation results in poor edge localization and multiscale edge representations cannot always separate salient contour features from irrelevant textural clutter. To overcome these problems, a novel algorithm is presented which first creates a multiscale edge representation using the Mallat wavelet transform and then recombines the edge map at each scale to create a single contour map where textural clutter has been minimized. This algorithm is then applied to natural and synthetic images containing contour features at different spatial scales and texture of varying spatial frequency and orientation. The results show that contour and textural features can be discriminated at each scale and the resulting contour map serves as a more effective representation on which subsequent localization and recognition tasks are based.