Adaptive Propagation Algorithm

To initially test the algorithm’s performance it was implemented into the JM 12.2 Reference Encoder [130]. The number of transitions when each macroblock was propagated vertically, horizontally and adaptively based on the 4x4 intra prediction algorithm described was then measured. The results given are the average across the first 100 frames of each sequence

Figure 6.8 shows the percentage reduction in transitions when the adaptive propagation algorithm is used compared to when both fixed horizontal, and fixed vertical, propagation are used. For this experiment, the sequences used were encoded using a constant quantisation parameter (see section 2.1.2) of 30. For all sequences the adaptive propagation algorithm reduces the number of tran-sitions. The extent of the reduction is to a large degree, sequence dependent.

The largest reduction in transitions occurs for sequences which contain a mixture of horizontally and vertically correlated macroblocks, such as hall monitor, paris and office. For these sequences, adaptively changing the propagation direction on a per macroblock basis provides a significant reduction in transitions compared to when either fixed horizontal or fixed vertical propagation is used. For other sequences where the majority of macroblocks are either vertically or horizontally correlated, such as riverraft and suzie, the adaptive propagation algorithm only provides a small reduction in transitions compared to using the best fixed prop-agation direction. However, in both cases the adaptive algorithm offers a signifi-cant reduction compared to when the worst fixed propagation direction is used.

For the rest of the analysis the adaptive algorithm is compared only to fixed horizontal propagation, the best fixed propagation direction for the sequences

Figure 6.8: Percentage reduction in transitions using adaptive propagation com-pared to fixed horizontal and fixed vertical propagation

Figure 6.9: Percentage reduction in transitions compared to horizontal propaga-tion when the 4x4 and 16x16 intra predicpropaga-tion results are used

studied.

Figure 6.9 compares the transition reduction achieved using the 4x4 intra pre-diction results and 16x16 intra prepre-diction results. A quantisation parameter of 30 was again used to encode the images. Apart from two sequences, missamerica and hall monitor, using the 4x4 intra prediction results within the adaptive algo-rithm causes a larger reduction in transitions. More significantly for the football, table and riverraft sequences using the 16x16 intra prediction results actually causes an increase in transitions. The worst case being the table sequence. For this sequence using the 16x16 intra prediction results causes a 1.5% increase in transitions compared to when fixed horizontal propagation is used.

Figure 6.10 shows the effect varying the quantisation parameter has on the percentage reduction in transitions achieved by the proposed algorithm. The percentage reduction in transitions does increase as the quantisation parameter increases. At higher quantisation levels, this effect is offset by the inaccuracy

Figure 6.10: Percentage reduction in transitions compared to horizontal propa-gation for a range quantisation parameters

of the intra prediction results. The intra prediction process is performed on unencoded macroblocks. However, it is encoded macroblocks which are saved to and loaded from external memory. The proposed algorithm implicitly assumes the unencoded and encoded macroblocks will be similar. When a high quantisation parameter is used there is the potential for this assumption not to hold and an incorrect propagation decision to be made. For sequences where the adaptive algorithm does not offer an inherent benefit, such as riverraft, this effect can result in the adaptive algorithm increasing transitions. This only occurs at when the quantisation parameter is set to a very high value, and consequently the encoded images are of very low quality.

The effect of using the proposed algorithm with DBM/VBM encoding and bus invert encoding is shown in figures 6.11 and 6.12. For these experiments a

Figure 6.11: Percentage reduction in transitions (compared to fixed horizontal propagation), using partial bus invert encoding with and without the proposed adaptive propagation algorithm

constant quantisation parameter of 30 was used. The results show that using the proposed algorithm in combination with either partial bus invert or DBM/VBM encoding provides a greater reduction in transitions than using bus invert or DBM/VBM encoding alone. The performance improvement is most notable when partial bus invert encoding is used. Using the adaptive propagation algorithm with bus invert encoding more than doubles the transition reduction achieved for sequences where the adaptive propagation algorithm alone offers a signifi-cant benefit. Despite this improvement, partial bus invert encoding still offers a smaller reduction in transitions than DBM/VBM encoding.

The performance improvement provided by the adaptive propagation algo-rithm is less notable when DBM/VBM encoding is used. With DBM/VBM encoding the number of transitions only decreases if the difference between con-secutive input values is increased by set amounts, dependent on the size of the

Figure 6.12: Percentage reduction in transitions(compared to fixed horizontal propagation) using DBM and 8-bit VBM encoding with and without the proposed adaptive propagation algorithm

VBM encoder used. Thus when an 8-bit VBM encoder is used, the impact of in-creasing the lag1-correlation using the proposed adaptive propagation algorithm is limited. However in the best case, using adaptive propagation still provides an additional 10% reduction in transitions. If a 4 bit VBM encoder is used the impact of adaptive propagation is greater as shown in Figure 6.13. The adaptive propagation algorithm providing upto a 15% reduction in transitions.