Transform coding of digitized image signals

A mathematical analysis of transform coding of digitized image signals is presented, with emphasis on two alternative types of transform: the discrete cosine transform (DCT) and the M-Hadamard transform (MHT). DCT is based on the matching of distortion to the bandpass character of the human visual system's contrast sensitivity. Because the power of high-frequency image patterns in natural scenes is smaller than those of low-frequency, matching can be accomplished using a power criterion. MHT coding is designed to realize constant image quality by using codewords of various lengths (between 1.8 and 2.9 bits per pixel). The codewords are classified according to the amount of detail they contain, and the quantized coefficients are presented in the gray code. An optimum uniform quantization of the coefficients defines the smallest quantizer step needed to insure the invisibility of reconstruction errors in flat image regions. One prospective area of application for these techniques is in high-volume image transmission. It is pointed out, however, that the data compression in the systems would have to be improved by including the dimension of time.