ISO/IEC 10918-1 : 1993(E)
Figure 8 3-sample prediction neighbourhood
FIGURE 8 [D08] 5 cm = 195 %
This encoding process may also be used in a slightly modified way, whereby the precision of the input samples is reduced
by one or more bits prior to the lossless coding. This achieves higher compression than the lossless process (but lower
compression than the DCT-based processes for equivalent visual fidelity), and limits the reconstructed image's worst-case
sample error to the amount of input precision reduction.
Modes of operation
There are four distinct modes of operation under which the various coding processes are defined: sequential
DCT-based, progressive DCT-based, lossless, and hierarchical. (Implementations are not required to provide all of
these.) The lossless mode of operation was described in 4.4. The other modes of operation are compared as follows.
For the sequential DCT-based mode, 8
8 sample blocks are typically input block by block from left to right, and block-
row by block-row from top to bottom. After a block has been transformed by the forward DCT, quantized and prepared for
entropy encoding, all 64 of its quantized DCT coefficients can be immediately entropy encoded and output as part of the
compressed image data (as was described in 4.3), thereby minimizing coefficient storage requirements.
For the progressive DCT-based mode, 8
8 blocks are also typically encoded in the same order, but in multiple scans
through the image. This is accomplished by adding an image-sized coefficient memory buffer (not shown in Figure 4)
between the quantizer and the entropy encoder. As each block is transformed by the forward DCT and quantized, its
coefficients are stored in the buffer. The DCT coefficients in the buffer are then partially encoded in each of multiple
scans. The typical sequence of image presentation at the output of the decoder for sequential versus progressive modes of
operation is shown in Figure 9.
There are two procedures by which the quantized coefficients in the buffer may be partially encoded within a scan. First,
only a specified band of coefficients from the zig-zag sequence need be encoded. This procedure is called spectral
selection, because each band typically contains coefficients which occupy a lower or higher part of the frequency spectrum
for that 8
8 block. Secondly, the coefficients within the current band need not be encoded to their full (quantized)
accuracy within each scan. Upon a coefficient's first encoding, a specified number of most significant bits is encoded first.
In subsequent scans, the less significant bits are then encoded. This procedure is called successive approximation. Either
procedure may be used separately, or they may be mixed in flexible combinations.
In hierarchical mode, an image is encoded as a sequence of frames. These frames provide reference reconstructed
components which are usually needed for prediction in subsequent frames. Except for the first frame for a given
component, differential frames encode the difference between source components and reference reconstructed
components. The coding of the differences may be done using only DCT-based processes, only lossless processes, or
DCT-based processes with a final lossless process for each component. Downsampling and upsampling filters may be
used to provide a pyramid of spatial resolutions as shown in Figure 10. Alternatively, the hierarchical mode can be used to
improve the quality of the reconstructed components at a given spatial resolution.
Hierarchical mode offers a progressive presentation similar to the progressive DCT-based mode but is useful in
environments which have multi-resolution requirements. Hierarchical mode also offers the capability of progressive
coding to a final lossless stage.
CCITT Rec. T.81 (1992 E)