Inter-frame subband video coding

Three inter-frame subband video coding models are studied. The three models, referred to as models I, II and III, are studied by employing three different inter-frame predictive coding methods together with subband analysis/synthesis using separable twodimensional quadrature mirror filter (QMF) banks. Models I and 11 use block matching and pel-recursive motion compensation whereas model III is based on a simple subband inter-frame difference coding scheme. Model I applies motion compensation to the full band image followed by subband analysis of the motion compensated prediction error. Model n applies subband analysis to the full band image and then uses motion compensated prediction in the resulting subbands. Uniform symmetrical quantisers are employed in each of the models to quantise the prediction error. The simulation results show good performance of the models, in terms of the peak signal-to-noise ratio (PSNR) and the entropy of the prediction error. Model I shows no significant difference in performance with either block matching or pel-recursive motion compensation. The results show that, for the test image sequence used, the performance of model II is significantiy better than model I and III. This effect is noted in simulations using both the block matching and pel-recursive motion compensation schemes. The results further indicate that Model II can be simplified by limiting the motion estimation to one subband and making available the same motion vectors to the other subbands, without compromising its performance. However, the PSNR results also suggest that the pel-recursive motion estimation, which is sensitive to large spatial gradients, works better in model II when the motion estimates are derived from the lowpass subbands. Model III, which encodes the subband frame differences, with no motion compensation, is shown to perform significantiy better than model I on a test TV sequence where the movement is small.