QUANTITATIVE COMPARISON OF TENSORIAL IMAGE DESCRIPTIONS FOR THE APPLICATION TO PERCEPTUAL GROUPING BY THE TENSOR VOTING TECHNIQUE

摘要

We have applied the perceptual grouping method known as tensor voting to grey-level images by using local orientation tensors. For that purpose, three tensorial description approaches are compared: Two methods are based on quadrature filters, namely Gabor filters and polar separable lognormal filters as introduced by Granlund and Knutsson. The third method employs structure tensors. The approaches are quantitatively evaluated on a set of test images with regards to their accuracy and noise robustness. According to the results, the tensors computed from a Gabor filter set show the highest angular precision at edges while simultaneously representing junctions correctly.