Wavelet transforms that map integers to integers (Q1271628)

The wavelet transforms described in this paper do not use integer arithmetic, but the floating point filtering operations are arranged in such a way that they map integer data to integer wavelet transforms and they are still invertible. Two approaches to solve the problem are described. The first is an adaptation of the precoder of Laroia, Tretter, and Farvardin, used in information transmission. The non-integers appear in the Haar transform because of a division by 2. These factors can be removed by using expansion coefficients for the filters. A more complicated but similar idea is used for more general wavelet filters. The result of the analysis is that expansion coefficients are used for high and low-pass filters and that the integer data are changed to compensate for the non-integer mapping of the filters, so that the result is again integer. A second approach is conceptually much simpler and is based on the lifting scheme for wavelet transforms. The decomposition into lifting steps allows for an easy and lossless inversion, even if the lifting steps are rounded to integers. The two approaches are essentially different and need not lead to the same algorithms. Several examples are given and they are applied to a set of test images to be compared with other transforms and coding methods given in the literature. Similar results were obtained in [\textit{S. Dewitte} and \textit{J. Cornelis}, IEEE Signal Proc. Lett. 4, No. 6, 158-160 (1997)] and [\textit{H. Chao}, \textit{P. Fisher} and \textit{Z. Hua}, Lect. Notes Pure Appl. Math. 202, 19-38 (1998)]. For a software implementation, see for example [\textit{G. Uytterhoeven} et al., ``WAILI: A software library for image processing using integer wavelet transforms'', In `Medical Imaging 1998: Image Processing', K. M. Hanson (ed.), SPIE Proceedings, Vol. 3338, International Society for Optical Engineering, 1490-1501 (1998)].