Fast integer-to-integer reversible lifting transform with reduced memory consumption

This paper addresses the problem of reducing the memory usage in the implementation of a reversible two-dimensional wavelet transform for image processing. In particular, we take a line-based approach by using a recursive algorithm to ease the synchronization among different buffer levels. In addition, since the reversible transform is non-linear, to preserve reversibility, we have to consider the order of the horizontal and vertical transforms in which the two-dimensional forward and inverse wavelet transform are decomposed. The proposed algorithm is suitable for integer-only devices (such as many FPGAs), reducing in more than 250 times the memory requirements, for a 5-Megapixel image with the well-known B5/3 wavelet transform. Moreover, it is several times faster (up to ten times) in cache-based systems, due to the better use of the cache memory if compared with the regular wavelet transform