Selective prediction error transmission using motion information

Hanssens, E Legat, Jean-Didier [UCL] Chupeau, B Macq, Benoît [UCL]

A new method to encode the prediction errors in a hybrid coding scheme is presented. The technique tries to reduce camera noise transcription in the prediction errors, thereby decreasing bit-rate. Camera noise is responsible for vector field inhomogeneity in regions of the image where the true motion is uniform. Vector field smoothing techniques, which cancel the effect of camera noise on motion vectors determination, form the central steps of the motion estimation algorithms used here. Nevertheless, these techniques do not always yield bit-rate decrease in DCT-based schemes. Indeed, although they decrease the camera noise impact on the estimated motion field by removing it from the estimation, this noise is reintroduced in the prediction errors, thereby increasing the bit-rate. To avoid this effect, prediction errors encoding is done by using information from the vector field. Prediction errors of regions where motion vectors are correctly estimated are coarsely quantized, while in other regions, a lower quantization factor is adopted. This technique yields a decrease of the P-frames bit-rate up to 40% as well as a PSNR decrease, but the image visual quality remains almost constant. (C) 1998 Elsevier Science B.V. All rights reserved.