MODIFIED INTRA PREDICTION SCHEME FOR H.264/AVC
3.3 Proposed Modified DC Prediction for Intra 4 × 4 Block
Fig. 3.1 Intra 4×4 prediction modes.
plane prediction compose all the four prediction modes for 16×16 luma block. The Intra 4×4 prediction is suitable for the parts with significant details, while the Intra 16×16 is applied to the smoother areas. The encoder chooses the best prediction mode to minimize the Lagrangian cost function, which takes both distortion and bit rate into consideration. Fig. 3.1 illustrates 16 samples of 4×4 block (labeled as a-p) which are predicted by previous-decoded samples in the neighboring blocks (labeled as A-M), when using the Intra 4×4 prediction.
For vertical or horizontal modes, the pixel values are extrapolated by upper samples or left samples, while other directional modes utilize the linear weighted average of reference sam- ples. For the DC mode, all the predicted pixels are formed by means of upper and left samples.
In Intra 4×4 prediction, each block can be predicted using either the DC mode or one of the eight directional modes. As we know, these eight directional modes are used to predict the regions with unified orientations. Therefore, it is better if the remainder, DC mode, can be used to predict some areas that the textures have no unified orientation. However, using DC mode to predict such areas is not very accurate because it uses one value to predict all pixels in the block, which can not show any kind of variations between them. Since the Intra 4×4 DC mode cannot provide accurate prediction for some areas without unified directions, we replace DC mode with MDCP mode, which is much simpler and efficient.
3.3 Proposed Modified DC Prediction for Intra 4 × 4 Block
In H.264/AVC, DC mode is used to predict regions with no unified direction and the predicted values of all pixels are same. But the correlation that exists between predicted pixels and reference pixels are not absolutely considered to predict the DC prediction mode. Thus, the prediction signal generated by DC prediction is not well matched to the original signal, and a large number of bits are required for encoding the difference between the predicted
and original signal. It is well known that Gaussian-like distribution can approximate local intensity variations in smooth image region. Original DC prediction always considers the ref- erence pixels are belong to upper and left reconstructed blocks. But the correlation between neighboring pixels would be attenuated while the distance is increased and negligible when pixels are far apart. Therefore, prediction accuracy is degraded if all of the pixels of a 4×4 block are predicted from upper and left blocks.
Statistics show that the vertical and horizontal predictions are more frequently used than other modes, implying higher correlations between the reference samples and the pixels to be predicted in these two directions [57]. Therefore, it is possible to enhance the intra prediction accuracy by employing more such directional predictions, which inspires to develop MDCP to replace DC mode in Intra 4×4, via combining vertical and horizontal predictions.
Furthermore, because only the upper and left reference samples are available, block is divided into three parts: the diagonal, upper right part and lower left part. Since the diagonal pixels have the equal distance between the upper and left reference, they can be predicted by both the corresponding upper and left reference samples. However, the upper right pixels are closer to the upper reference while the lower left pixels closer to the left ones. Therefore, we mainly use the upper reference samples as the major component for the upper right part prediction while using the left reference samples for the lower left part.
In other words, we use the vertical prediction as the major component for the upper right part, while employing the horizontal prediction as the major component for the lower left part.
Since correlation between pixels exists in both vertical and horizontal direction, we cannot only use the vertical or horizontal prediction directly in a 4×4 block. As a result, when we predict the upper right part, we use the left predicted pixel to adjust the vertical prediction, while using the upper predicted pixel as an adjuster to the horizontal prediction for the lower left part similar to [58].
As illustrated in Fig. 3.2, Pi,j (0≤i,0≤ j)denotes the predicted value of the pixel in the ith row and jth column of current 4×4 block.UjandLi denote the reference samples reconstructed from upper and left blocks respectively. In order to make a weighted prediction, a larger coefficient 3/4 is assigned to the major reference samples while a smaller coefficient 1/4 to the adjustment ones.
The denominators 4 make the division operation to be implemented by shift operation.
Pi,jcan be calculated as:
3.3 Proposed Modified DC Prediction for Intra 4×4 Block 47
Fig. 3.2 Illustration of Intra 4×4 prediction.
P(i,j)=
3(Uj+Pi,j−1+2)>>2= [(Uj<<1) +Uj+Pi,j−1+2]>>2 ifi< j;
(Li+Uj+1)>>1 ifi= j;
3(Lj+Pi,j−1+2)>>2= [(Lj<<1) +Lj+Pi,j−1+2]>>2 ifi> j.
(3.1)
Since some of the predicted values are based on the other predicted values, we have to compute the predicted values in the following order. The diagonal pixels should be computed first, which are only based onUjandLi. Then, the upper right and lower left parts can be computed. The modified expressions of MDCP are listed as follows. It can be seen that the longer the distance between the reference sample and the pixel is, the smaller the coefficient becomes.
P0,0= 1 2U0+1
2L0 (3.2)
P1,1= 1 2U1+1
2L1 (3.3)
P2,2= 1 2U2+1
2L2 (3.4)
P3,3= 1 2U3+1
2L3 (3.5)
P0,1= 3 4U1+1
4P0,0=3
4U1+1
8(U0+L0) (3.6)
P0,2= 3 4U2+1
4P0,1=3
4U2+ ( 3
16U1+ 1
32U0+ 1
32L0) (3.7)
P1,2= 3 4U2+1
4P1,1=3
4U2+1
8(U1+L1) (3.8)
P0,3= 3
4U3+1
4P0,2= 3
4U3+ ( 3
16U2+ 3
64U1+ 1 128U0+
128L0) (3.9) P1,3= 3
4U3+1/4P1,2= 3
4U3+ ( 3
16U2+ 1
32U1+3
4L1) (3.10)
P2,3= 3 4U3+1
4P2,2=3
4U3+1
8(U2+L2) (3.11)
P1,0=3 4L1+1
4P0,0=3 4L1+1
8(U0+L0) (3.12)
P2,0=3 4L2+1
4P1,0=3
4L2+ ( 3
16L1+ 1
32U0+ 1
32L0) (3.13)
P2,1=3 4L2+1
4P1,1=3 4L2+1
8(U1+L1) (3.14)
P3,0=3 4L3+1
4P2,0=3
4L3+ ( 3
16L2+ 3
64L1+ 1
128U0+ 1
128L0) (3.15) P3,1=3
4L3+1
4P2,1=3
4L3+ ( 3
16L2+ 1
32U1+ 1
32L1) (3.16)
P3,2=3 4L3+1
4P2,2=3 4L3+1
8(U2+L2) (3.17)
We also compare the computation complexity between iDWP and MDCP. The compu- tation complexity of iDWP to predict one pixel includes 1 multiplication, 2 additions, 1