Inter and Inter-Layer Residual Prediction Data Reuse Method

In the Inter-layer prediction, the information including motion information, texture, and residuals are up-sampled from the base layers and used for the prediction in enhancement layers. From our observation, we found that some common parts of data and computation can

53

be reused during the Inter-layer prediction process. Therefore, in this subsection, we first analyze the matching criteria of different prediction modes to find out the common parts and distinct parts. Afterwards, based on our analysis, we propose our data reuse method for Inter and Inter-layer prediction.

3.2.1. Matching Criteria of Different Prediction Modes

To evaluate which prediction mode should be selected as best mode or which position should be chosen as the best matching position, the rate distortion cost is widely adopted in most of video coding standards to achieve best tradeoff between coding rate and distortion. The rate distortion cost (RDCost) can be expressed as follows.

(3-1) where λ is the Lagrange multiplier, R refers to the rate for coding the motion information, and D is the distortion which can be defined as follows.

(3-2) where C and F are the current and reference pixels, respectively.

However, for different prediction modes, the calculation of distortion term D has slight difference. The distortion term of Inter prediction DInter can be calculated as follows.

(3-3) For Inter-layer residual prediction, the distortion term DILR can be calculated as follows.

(3-4) where B is the residuals up-sampled from base layer.

For Inter-layer motion prediction, the distortion term DILM can be calculated as follows.

(3-5) For Inter-layer motion + residual prediction, the distortion term DILMR can be calculated as

54

follows.

(3-6) From above distortion terms calculation, we can observe some properties. First, for the Inter and Inter-layer motion prediction mode, the only difference between DInter and DILM

calculation is only on that the motion vector predictor. However, although only slight difference has been observed from the equations of DInter and DILM, the reference data is quite different actually between these two prediction modes since different motion vector predictors will acquire different reference data for the search process. Therefore, there is nothing except current data can be reused when computing the DInter and DILM. Second, for the Inter and Inter-layer residual prediction modes, we can observe that the only difference between these two prediction modes is that the residual information should be subtracted from the current data ahead as Fig. 3.8 shown. In other words, the current and reference data can be reused for both of Inter and Inter-layer residual prediction modes. Similarly, the same situation can also be seen from the Inter-layer motion and Inter-layer motion + residual prediction modes as Fig. 3.9 shown. As a result, through the reusing of current and reference data for different prediction modes, the data bandwidth requirements and computational components can be reduced significantly.

IME module

Fig. 3.8. Illustration of distortion terms calculation for (a) Inter prediction : DInter and (b) Inter-layer residual prediction : DILR

55

Fig. 3.9. Illustration of distortion terms calculation for (a) Inter-layer motion prediction : DILM and (b) Inter-layer motion + residual prediction : DILMR

3.2.2. Proposed Data Reuse Method

Fig. 3.10 exhibits our proposed data reuse scheme for different prediction modes. Compared to the original form of distortion terms calculation, we only slightly change the subtraction order of distortion terms calculation in our proposal. The benefits of our proposal are described as follows. First, the computational modules can be shared in hardware implementation. In the original form, two hardware modules may be implemented to generate the distortion terms of DILR and DInter or DILMR and DILM intuitively. However, through our proposed data reuse scheme, the distortion term of one prediction mode can be generated first and the other one will be generated right after so that the unnecessary repeated computations can be avoided. For example, through our proposal, the distortion term of DILR can be calculated following the calculation of DInter. Hence, the repeated subtractions for current and reference data can be saved. Second, the reference data can be reused for different prediction modes. More precisely, the reference data for Inter mode prediction can be reused for the Inter-layer residual prediction. Similarly, the reference data for Inter-layer motion prediction can be reused for the Inter-layer motion + residual prediction as well. As a result, the distortion terms calculation can be redefined as follows when applying our proposed data reuse scheme.

56

For Inter-layer residual prediction, the distortion term DILR can be calculated as follows.

(3-7) For Inter-layer motion + residual prediction, the distortion term DILMR can be calculated as follows.

(3-8)

Reference MB pixels

Inter-layer residual module Search

area

Current MB pixels

Base layer residuals

DILR DInter

(x,y)

(a)

Reference MB pixels

Search area

Current MB pixels

Base layer residuals

D_ILMR DILM

(x`,y`)

Inter-layer motion + residual module

(b)

Fig. 3.10. Proposed data reuse scheme for (a) Inter and Inter-layer residual prediction and (b) Inter-layer motion and Inter-layer motion + residual prediction

3.2.3. Results

Since our proposed data reuse scheme is lossless approach, there is no rate distortion performance difference between our proposal and original form. However, for the data

57

bandwidth savings, our proposal can achieve 44.45% data bandwidth savings per macroblock when search range of ±8 is applied.

3.3. Low Bandwidth Inter and Inter-Layer Motion Prediction