Fast multiscale directional filter bank with applications to texture retrieval

The contourlet transform has emerged recently as an efficient directional multiresolution image decomposition. In order to enhance its distinguishing power for texture images, the high frequency band is split by half and results in the multiscale directional filter bank (MDFB). However, this raises the computational complexity considerably because an extra set of scale and directional decomposition is required to perform on the full image size. In this paper, we develop a new framework for the MDFB, in which directional decomposition is performed prior to the scale decomposition in the first two scales. Experiments show that 30.6%-33.2% saving in computational time can be achieved. Meanwhile, it has the same flexibility in frequency partitioning and similar texture retrieval performance as the original algorithm of the MDFB.