Boundary detection in complex wavelet domain with edge flow

A novel boundary detection method based on multiscale edge flow is proposed, which integrates the local edge flow field into a dual tree complex wavelet transform (DT-CWT) and takes into account the boundary pixel dependence across scales. The shift-invariance of DT-CWT and the use of predictive edge flow vector propagation help assure the good localization of boundaries; the tracking of boundaries from coarse scale to fine scale and the directional sparse image can reject the noise. The new method computes both the intrascale predictive relationship and interscale dependence in a geometric way. Experimental results on natural and carton images validate its usefulness particularly on geometric images with weak boundaries and noise.