10.6 Texture coding

10.6 Texture coding

The intra VOPs and motion compensated inter VOPs are coded with 8 × 8 block DCT. The DCT is performed separately for each of the luminance and chrominance planes.

For an arbitrarily shaped VOP, the macroblocks that completely reside inside the VOP shape are coded with a technique identical to H.263. Boundary macroblocks that are of intra type are padded with horizontal and vertical repetition. For inter macroblocks, not only is the macroblock repeatedly padded, but also the region outside the VOP within the block is padded with zeros. Transparent blocks are skipped and therefore are not coded. These blocks are then coded in a manner identical to the interior blocks. Blocks that lie outside the original shape are padded with 128, 128 and 128 for the luminance and the two chrominances in the case of intra and 0, 128 and 128 for inter macroblocks. Blocks that belong neither to the original nor the coded arbitrary shape but to the inside of the bounding box of the VOP are not coded at all.

10.6.1 Shape adaptive DCT

At the boundary macroblocks, the horizontally/vertically padded blocks can be coded with a standard 8×8 block DCT. This padding removes any abrupt transitions within a block, and hence reduces the number of significant DCT coefficients. At the decoder the added pixels are removed by the help of shape parameters from the decoded BABs.

Since the number of opaque pixels in the 8 x 8 blocks of some of the boundary macroblocks is usually less than 64 pixels, it would have been more efficient if these opaque pixels could have been DCT coded without padding. This method of DCT coding is called shape adaptive DCT (SA-DCT). The internal processing of SA-DCT is controlled by the shape information that has to be derived from the decoded BAB. Hence only opaque pixels within the boundary blocks are actually coded. As a consequence, in contrast to standard DCT, the number of DCT coefficients in an SA-DCT is equal to the number of opaque pixels.

There are two types of shape adaptive DCT, one used for inter blocks, known as SA-DCT and the other for intra blocks, known as ∆ SA-DCT, which is an extension of SA-DCT. For both cases, a two-dimensional separable DCT with a varying length of basis vector is used.

The basic concept of SA-DCT is shown in Figure 10.21. Segments of the opaque pixels are encapsulated in the picture grids of 8 x 8 pixel blocks as shown in Figure 10.21a. Each row of the pixels is shifted and aligned to the left, as shown in Figure 10.21b. The aligned pixels are then one-dimensionally DCT coded in the horizontal direction with variable basis functions, where the lengths of the basis functions are determined by the number of pixels in each line. For example, the first pixel of the segment is represented by itself, as a DC coefficient.

Figure 10.21: An example of SA-DCT (a original segment) (b ordering of pixels and horizontal SA-DCT) (c location 1D coefficients) (d location of samples prior to vertical SA-DCT) (e ordering of 1D samples and vertical SA-DCT) (f location of 2D SA-DCT coefficients)

The second line of the pixels is DCT coded with a three-point transform, and the third line with a five-point transform, and so on. The coefficients of the N-point DCT, c_j, are defined by:

(10.12) 

and

Figure 10.21c illustrates the horizontal DCT coefficients, where the DC values are represented with a dot. These coefficients now become the input to the second stage one-dimensional DCT in the vertical direction. Again they are shifted upwards and aligned to the upper border, and the N-point DCT is applied to each vertical column, as shown in Figure 10.21e. The final two-dimensional DCT coefficients are shown in Figure 10.21f. Note that, since the shape information from the decoded BABs is known, these processes of shifting and alignments in the horizontal and vertical directions are reversible.

The ∆SA-DCT algorithm that is used for intra coded macroblocks is similar to SA-DCT but with extra processing. This additional processing is simply calculating the mean of the opaque pixels in the block, and subtracting the mean from each individual pixel. The resultant zero mean opaque pixels are then SA-DCT coded. The mean value is separately transmitted as the DC coefficient.

Zigzag scanning, quantisation and the variable length coding of the SA-DCT coefficients operations are similar to those of standard DCT used in H.263.