Nonlinear neural prediction in 1D DPCM for efficient image data coding

Neural net architectures, with a hidden layer or functional links have been utilized to generate predictions for 1D differential pulse-code modulation (DPCM) applied to still image coding. In this approach, the predictor is designed by supervised training based on a typical sequence of pixel values, i.e. the values of the coefficients of the predictor are determined by training on examples. Nonlinear and linear correlations are exploited. Computer simulation experiments have been carried out to evaluate the resulting performance. At a transmission rate of 1 bit/pixel, for the images LENA and BABOON, the 1D neural network DPCM provides a 4.17 and 3.74 db improvement in peak SNR, respectively, over the standard linear DPCM system