Page 104 - Video Coding for Mobile Communications Efficiency, Complexity, and Resilience

Section 3.5. The MPEG-4 Standard 81

3.5.4 Texture Coding
For I-VOPs, texture refers to the luma and chroma values (i.e., the video signal).
For motion-compensated VOPs, texture refers to the luma and chroma residual
errors remaining after motion compensation (i.e., the DFD signal). The process
of texture coding involves the following steps: padding, DCT, quantization,
INTRA coe,cient prediction, scanning, and variable-length encoding.

3.5.4.1 Padding
Like H.263 and most other video coding standards, MPEG-4 encodes texture
information using a block-based 8 × 8 DCT. In this process, internal MBs are
encoded directly, whereas boundary MBs must /rst be padded. The aim of this
padding process is to remove abrupt transitions within the macroblock, thus
reducing the number of signi/cant DCT coe,cients. Note that during texture
coding, exterior MBs are not coded.
For motion-compensated boundary MBs, pels outside the VOP are padded
with zero. For INTRA boundary MBs, pels outside the VOP are padded using
the following low-pass extrapolation (LPE) procedure:
1. Calculate the mean value of the macroblock pels that lie within the
VOP. Use this value for padding the macroblock pels that lie outside
the VOP.
2. Starting at the top-left corner of the macroblock, proceed in scanning
order to the bottom-right corner, replacing each pel f(x; y) that lies
outside the VOP with the average value of its four neighbors; i.e.,
f(x; y)=(f(x − 1;y)+ f(x +1;y)+(f(x; y − 1) + f(x; y + 1))=4. The
neighboring pels should lie within the VOP; otherwise they are not
considered in the averaging process and the equation is modi/ed ac-
cordingly.

3.5.4.2 DCT
The internal MBs and the padded boundary MBs are then transformed using
a 2-D 8 × 8 forward DCT.

3.5.4.3 Quantization

The resulting DCT coe,cients are quantized using one of two methods. The
/rst method is very similar to H.263 quantization and uses a /xed quantization
step size for the whole macroblock. The second method, however, uses one of
two default quantization matrices (or scaled versions of them) to modify the