Neural network approach to DPCM system design for image coding

Abstract

This paper presents a neural network approach to differential pulse code modulation (DPCM) design for the encoding of images. Instead of traditional algorithms for the computation of the relevant coefficients, such as the autocovariance and autocorrelation methods, the predictor is designed by supervised training of a neural network on examples, i.e. on a typical sequence of pixel values. This allows the use of nonlinear as well as linear correlations. Efficient and fast neural net architectures, for nonlinear one-dimensional DPCM (NNDPCM) as well as two-dimensional adaptive DPCM (NNADPCM), have been designed and applied to still image coding. Computer simulation experiments have shown that the resulting encoders work very well. At a transmission rate of 1 bitpixel, the 1-D NNDPCM offers an advantage of about 4 dB in peak signal-to-noise ratio over the standard linear DPCM system. At a bit rate of 0.525 bit/pixel, the 2-D NNADPCM achieves 29.5 dB for the 512 × 512 Lena image, while there is little visible distortion in the reconstructed image. This performance is comparable to that of the best schemes known to date, whether DPCM based or not, while maintaining a lower encoding complexity. Furthermore, this establishes that there is substantial amount of nonlinear content available for 1-D and 2-D prediction in DPCM image coding.