Remapping of FGS Layers

In SVC [23], one FGS slice may include multiple FGS layers. Since the FGS layers are coded in a sequential manner, simply using the prioritized block coding in each FGS layer is not

Chapter 5. Applications in Scalable Video Coding Standard

Cycle 0 Cycle 1 Cycle 2 Cycle 3

S

The n-th (EOB, Run, Level) symbol coding in Block m R

(m,n) The refinement coding of n-th zigzag ordered coefficient in Block m

W Waiting Cycle

Figure 5.16: Prioritized block coding with the alternative assignment for Enable MB Coding flags.

sufficient to provide a ROI functionality over the entire bit rate range. For instance, if a FGS slice has two FGS layers, the FGS layer 1 must be coded prior to the FGS layer 2. As a result, the lower priority region in the FGS layer 1 is coded prior to the higher priority region in the FGS layer 2.

To break the boundaries of FGS layers, we employ a layer remapping technique. With the remapping of layers, the lower priority region in the FGS layer 1 can be coded in the FGS layer 2. For the remapping, we modify the syntax of code block pattern on a macroblock-by-macroblock basis. The code block pattern for the macroblock-by-macroblocks in the lower priority region is set to 0 while coding the FGS layer 1. In such a way, the macroblocks of lower priority will be quantized and coded in the FGS layer 2. For the FGS layer 1, the layer remapping is a technique for providing a binary ROI functionality. However, to maintain the scalability and provide a graceful ROI functionality, the prioritized block coding is still essential.

5.4.2.5 Preliminary Experiments

For demonstrating the ROI functionality, a preliminary experiment was conducted in [21], where we use Foreman sequence in CIF resolution as an example and assign the ROI partition according to the shape information. Figure 5.17 shows the setting of ROI partition. Particularly, in this example, the shifting factor of a macroblock is determined by the proportion of pixels

Figure 5.17: Partition of region-of-interest for graceful selective enhancement.

that are covered by the foreground region. In Figure 5.17, the foreground intensity shows the distribution of the shifting factors.

Figure 5.18 shows a regional PSNR comparison. As shown in the part (a) of Figure 5.18, enabling our prioritized block coding can dramatically increase the PSNR of the foreground re-gion. However, without the layer remapping, our scheme has no improvement at the boundaries of FGS layers. This is because there are no additional refinement coefficients for the foreground region due to the layer constraint. But, with the layer remapping, we can provide a ROI func-tionality over a wide range of bit rates. In particular, simply using the layer remapping is not sufficient to show a noticeable ROI functionality over the entire bit-rate range.

The part (b) of Figure 5.18 illustrates the PSNR comparison for the background region.

Since the foreground region is assigned with higher coding priority, the PSNR performance for the background region is the reciprocal case of the foreground region.

Figure 5.19 further compares the subjective quality between the cyclical block coding [26]

and our prioritized block coding with layer remapping. As shown, the proposed scheme con-sistently show better foreground quality at different bit rates. The ROI functionality is achieved over a wide range of bit rates.

To further assess the coding efficiency, we measure the increase of file size. Specifically, we find that enabling the ROI functionality has about 0.4%∼1.7% increase on file size. In other words, the ROI functionality is achieved with little changes on coding efficiency. In fact, the layer remapping may sometimes have slightly better coding efficiency because the significance map of the background region is only transmitted once in a FGS slice.

Chapter 5. Applications in Scalable Video Coding Standard

Foreground, GOP = 1

Bit rate (Kbits/s)

300 600 900 1200 1500 1800

PSNR-Y (dB)

29 30 31 32 33 34 35 36 37

Cyclical Block Coding (CBC) Prioritized Block Coding (PBC) CBC + Remapping

PBC+ Remapping

(a)

Background, GOP = 1

Bit rate (Kbits/s)

300 600 900 1200 1500 1800

PSNR-Y (dB)

28 29 30 31 32 33 34 35 36

Cyclical Block Coding (CBC) Prioritized Block Coding (PBC) CBC + Remapping

PBC + Remapping

(b)

Figure 5.18: Regional PSNR comparison. (a) Comparison of foreground region. (b) Compari-son of background region.

(a) Cyclical Block Coding (b) Prioritized Block Coding + Remapping

at 300kbits/s at 300kbits/s

(c) Cyclical Block Coding (d) Prioritized Block Coding + Remapping

at 700kbits/s at 700kbits/s

(e) Cyclical Block Coding (f) Prioritized Block Coding + Remapping

at 1100kbits/s at 1100kbits/s

Figure 5.19: Subjective quality comparison between cyclical block coding and the priotrized block coding with layer remapping.

Chapter 5. Applications in Scalable Video Coding Standard

5.5 Summary

In this chapter, we give a brief introduction to the emerging scalable video coding (SVC) stan-dard [23]. Moreover, we illustrate the applications of our EMFGS and SBR schemes in the SVC framework. Although the EMFGS and SBR are mainly designed for MPEG-4 FGS [15], we have shown that these techniques can also be extended and tailored for the SVC standard [23].

Specifically, the proposed EMFGS can be employed for improving the coding efficiency of anchor frames. In SVC [23], the anchor frames and their enhancement-layer frames are coded in a way similar to that used by MPEG-4 FGS [15]. In Chapter 3, we have shown that using the enhancement layer for prediction can improve the coding efficiency. Therefore, the same technique used in EMFGS can be utilized for the prediction of anchor frames.

In addition, the idea of SBR can be extended for improving the coding efficiency of FGS layers. In SVC [23], the FGS layers are coded by a cyclical block coding scheme [26]. Each block is equally coded with one symbol in a coding cycle. Through the concept of SBR, a prioritized block coding is proposed to have the symbols with better rate-distortion performance be coded with higher priority. By reshuffling the symbols according to their stochastic rate-distortion performance, an improvement of 0.5dB is achieved.

Except for improving the coding efficiency, the prioritized block coding can also serve the ROI functionality. The macroblocks in the ROI can be coded prior to those that are outside the ROI. To achieve this, the ROI partition and priority information are explicitly coded in the bit-stream. To minimize the overhead, we employ an efficient priority assignment scheme.

Preliminary experiments have shown that our prioritized block coding can provide a graceful ROI functionality with little changes on coding efficiency.

Most of the works presented in this chapter are not fully developed. There are still lots of spaces for further improvements.

Conclusion

To serve multimedia applications under a heterogeneous environment, MPEG-4 has defined the Fine Granularity Scalability (FGS) [15], which provides a DCT-based scalable approach in a layered fashion. In Chapters 3 and 4, we have shown that current approach suffers from poor coding efficiency and subjective quality. For improving the coding efficiency, we propose an enhanced mode-adaptive FGS (EMFGS) algorithm and a context-adaptive bit-plane coding (CABIC) scheme. Moreover, to improve the subjective quality, we further develop a stochastic bit reshuffling (SBR) scheme based on the CABIC framework. In summary, as compared to the algorithm in MPEG-4 FGS [15], our schemes achieve a PSNR gain of 2∼3.5dB and provide better subjective quality. Although the proposed schemes are specifically developed for MPEG-4 FGS [15], in Chapter 5, we also show their applications in the upcoming scalable video coding standard [23]. In the following, we summarize the works in this thesis and show how the proposed schemes can be further improved.