Data compression using matching pursuits algorithms

摘要

A transform, preferably a frequency transform, is first applied to multi-dimensional data to be compressed and then a sequence of one-dimensional matching pursuits algorithms is used to code the transformed data. Specifically, the transformed data is convolved with a plurality of one-dimensional basis functions to generate a corresponding plurality of convolved data sets, the location of the convolved data set representative of greatest magnitude is determined in a first direction, the part of the transform data surrounding the location is represented by an 'atom' 22, the 'atom' is subtracted from the data set to create a new data set, and repeating these locating and subtracting steps. Atoms 26 may also be determined in a second direction, preferably orthogonal to the first, for subtraction. The invention finds application in video and still image coders, including real time coders and very low or scalable bit rate coders.