Processing for compression - The MPEG - page 98

Figure 3.68: In two's complement coding, sign extension takes place as the level falls. These sign-extended bits
are redundant and can be eliminated by multiplying by a level-dependent factor and neglecting the trailing zeros as
shown in (b).
Taking bits out of the middle of a word is not straightforward and in practice the solution is to multiply by a level-
dependent factor to eliminate the sign extension bits. If this is a power of 2 the useful bits will simply shift left up to
the sign bit. The right-hand zeros are simply omitted from the transmission. On decoding a compensating division
must be performed. The multiplication factor must be transmitted along with the compressed data so this can be
done. Clearly if only the sign extension bits are eliminated this process is lossless because exactly the same data
values are available at the decoder.
The reason for sub-band filtering and transform coding now becomes clear because in real signals the levels in
most sub-bands and the value of most coefficients is considerably less than the highest level.
In many cases the coding gain obtained in this way will not be enough and the wordlength has to be shortened
even more. Following the multiplication, a larger number of bits are rounded off from the least significant end of the
word. The result is that the same signal range is retained but it is expressed with less accuracy. It is as if the
original analog signal had been converted using fewer quantizing steps, hence the term requantizing.
During the decoding process an inverse quantizer will be employed to convert the compressed value back to its
original form. Figure 3.69 shows a requantizer and an inverse quantizer. The inverse quantizer must divide by the
same gain factor as was used in the compressor and re- insert trailing zeros up to the required wordlength.
Figure 3.69: Coding gain is obtained by shortening the sample or coefficient wordlength so fewer bits are needed.
The input values are scaled or amplified to near maximum amplitude prior to rounding off the low-order bits. The
scale factor must be transmitted to allow the process to be reversed at the decoder.
A non-uniform quantization process may be used in which the quantizing steps become larger as the signal
amplitude increases. As the quantizing steps are made larger, more noise will be suffered, but the noise is