Texture analysis and classification with quincunx and tree-structured wavelet transform

While using conventional two-dimensional wavelet transform for texture analysis and classification the image decomposition is carried out with separable filtering along the abscissa and ordinate using the same pyramidal algorithm as in the one-dimensional case. This process is simple and can be implemented easily in practical applications, however, it is rotation-sensitive and some information may be lost since the decomposition is performed only in low frequency channels. In this paper the quincunx transform using nonseparable sampling and filters is substituted for conventional dyadic transform. Since the energy of natural textures is mainly concentrated in the mid-frequencies, this transform can preserve more of the original signal energy and can provide more reliable description of the texture. At the same time, the tree-structured wavelet transform or wavelet packets is applied instead of using the pyramid-structured one. With this transform, we are able to zoom into any desired frequency channels for further decomposition and a series of subimages with the largest energy can be obtained for an image. In comparison with conventional wavelet transform, it can be concluded that this transform can still reach higher classification accuracy especially for the characterization of noisy data.