Directional wavelet bases constructions with dyadic quincunx subsampling (Q1645279): Difference between revisions

Tensor product (bi)orthogonal wavelet bases are widely known for their superior performance in image compression. However, tensor product wavelets lack directionality: they can only capture edge singularities along the horizontal or vertical directions. Therefore, being multiscale representation systems, wavelet bases and wavelet frames with directionality have been extensively investigated in the literature. This paper studies two-dimensional directional (bi)orthogonal wavelet bases using the dyadic dilation matrix for the refinable function and dyadic quincunx subsampling for the wavelet functions. This paper focuses on a particular configuration with six directions for the two-dimensional wavelet bases through filter bank construction. In particular, the authors establish the perfect reconstruction conditions in Theorem 1 for directional orthonormal wavelets and in Theorem 3 for directional biorthogonal wavelets. For the critically sampled case (i.e., non-redundant case), the paper shows in Section 4.2 that the Fourier transforms of such directional wavelet bases cannot be continuous, and consequently lack spatial localization. The authors show that this shortcoming can be remedied by adding a little bit redundancy, more precisely, directional wavelet frames with low redundancy can be achieved by proper modification and adaption techniques developed in this paper. Following the approach of two-dimensional hexagonal biorthogonal wavelets and using optimization techniques, the authors develop a numerical method to construct such two-dimensional directional biorthogonal wavelets in Section 5. Such directional wavelet bases studied and constructed in this paper are expected to be useful for image compression and other data processing tasks.