Modified bandlet transform with reward prediction of image geometry

Abstract

With emphasis on total variation being a reward prediction of image geometry, a modified bandlet transform is proposed. The new scheme is more conformed to human vision than the original one. It takes advantage of the sparse distribution of image geometry in the wavelet domain and retransforms wavelet coefficients in a top-to-bottom way. The point is that the searching space for optimal geometries can be refined in advance and the retransform can be constrained within the vicinity of the desired geometry except the ambient space, where it is assumed image geometries exist in all size of squares. Experimental results and complexity analysis validate the improvement of the proposed scheme especially in determination of the optimal image geometry through a decision-making strategy.