What DV standards call level is the magnitude – that is, the absolute value – of the AC coefficient. Sign is coded separately.

For a more elaborate description, and the quantization tables, see Symes, cited on page 535.

quality, and in any event users do not expect the same picture quality in shuttle mode as in normal playback.

DV quantization

The main challenge of DV encoding is to determine suitable quantization matrices for a segment’s AC coefficients, such that when all of the quantized coefficients are subject to variable-length encoding, the VLEcoded coefficients just neatly fit in the available space. The goal is to quantize the AC terms as finely as possible, without exceeding the capacity of a segment. In essence, this is a form of rate control. Quantization of a segment takes place after the DCT, using this algorithm:

•First, each block in the segment is assigned to a class from 0 (fine) to 3 (coarse), representing the block’s spatial complexity. DV standards provide a table that suggests how an encoder can assign a class number according to the magnitude of the largest AC term of a block; however, use of that table is not mandatory.

•Then, up to 15 trial quantization passes are made, to determine a quantization number (QNO) from 0 (coarse) to 15 (fine). Class number and quantization number are combined to determine a quantization matrix according to tables in the standard. For each trial QNO, DCT coefficients are quantized and zigzag scanned. Nonzero AC coefficient are identified; {run length, level} pairs are computed for each nonzero coefficient; and the required number of variable-length-encoded bits is accumulated. Quantization is eased by the fact that the entries in the quantization matrices are all powers of two (1, 2, 4, 8, 16); each coefficient quantization operation is merely a binary shift. The lookup and assembly of VLE-coded bitstream need not be performed at this stage; it suffices for now to accumulate the bit count.

The final QNO for the segment is the one that produces the largest number of bits not exceeding the capacity of the segment – for DV25, 500 bits for luma AC coefficients (including four 4-bit EOBs), and 340 bits for chroma AC coefficients (including two 4-bit EOBs). Once the segment’s QNO is determined, VLE coding

takes place, and the CMs of the segment are assembled. Each CM starts with one byte containing its QNO and error concealment status (STA) bits. Each block

This scheme is described as threepass; however, the first pass is trivial

includes its DC coefficient, its mode (8-8-DCT or 2-4-8-DCT), and its class. Finally, the VLE-coded AC coefficients are distributed in a deterministic three-pass algorithm – first to the associated block, then to unused space in other blocks of the same CM (if space is available), and finally to unused space in other CMs of the segment. QNO has been chosen such that sufficient space for all coefficients is guaranteed to be available: Every bit of every coefficient will be stored somewhere within the segment.

Each CM comprises 77 bytes, including by a 1-byte header. In DV25, a CM includes four coded luma blocks and two coded chroma blocks:

•A coded luma block totals 14 bytes, and includes a 9-bit DC term, one mode bit, and a 2-bit class number. One hundred bits are available for AC coefficients.

•A coded chroma block totals 10 bytes, and includes a 9-bit DC term, one mode bit, and a 2-bit class number. Sixty-eight bits are available for AC coefficients.

For 4:2:2 subsampling in DV50, a CM has four blocks, not six; space that in 4:1:1 or 4:2:0 would be allocated to luma blocks is available for overflow data. For 3:1:0 subsampling (used in SDL, to be described in a moment), a CM has eight blocks, not six: Each luma block has 10 bytes; each chroma block has 8 bytes.

DV digital interface (DIF)

The superblocks that I have mentioned form the basis for digital interface of DV bitstreams. A 3-byte ID is prepended to each 77-byte coded macroblock to form an 80-byte digital interface (DIF) block.

A coded DV25 superblock is represented by 135 video DIF blocks. That is augmented with several nonvideo DIF blocks to form a DIF sequence of 150 DIF blocks:

•1 header DIF block

•2 subcode DIF blocks

•3 VAUX DIF blocks

•9 audio DIF blocks

•135 video DIF blocks

Realtime DV25 video requires 10 or 12 DIF

sequences – that is, about 1500 or 1800 DIF blocks – per second. DV50, and DV100 systems have comparable structures, but different data rates.

DVCPRO and DVCAM data rate is 25 Mb/s, identical to consumer DV; however, the physical magnetic tape is more robust and more suitable for professional use.

DVCPRO50 DVTRs are standardized in SMPTE’s D-7 series; DV50 SD bitstreams are recorded onto 6 mm tape in DVC-style cassettes.

Panasonic introduced the DVCPRO HD DVTR format, which was subsequently standardized as SMPTE D-12. See page 482.

Once packaged in DIF sequences, DV bitstreams can be conveyed across the IEEE 1394 interface, also known as FireWire and i.LINK, that I described on page 167.

IEEE 1394 is suitable for consumer use, and is widely used in desktop video. For professional applications, DIF sequence bitstreams can be transported across various interfaces including the SDTI interface that

I introduced on page 441. (The 3-byte ID is unused in DV-over-SDTI.) DIF sequences can be stored in files – for example, QuickTime or MXF files.

Consumer DV recording

DV25 was widely adopted for consumer SD recording on MiniDV cassettes. Several schemes to extend DV to HD were described in the first edition of this book; none of these were commercialized. Consumer HD recording on MiniDV uses the HDV system, outlined on page 161.

Professional DV variants

DV technology was introduced for consumer videotape recording, and was widely deployed in consumer camcorders and desktop video editing systems. DV videotape technology was adapted to professional videotape recorders – first for SD as DVCPRO (D-7) and DVCAM, then for HD as DVCPRO HD (D-12). The DV compression system made the transition into products using hard disk drive and flash media.

DV50, also for SD, has twice the data rate, twice as many macroblocks per second (or per frame), and twice as many superblocks per second (or per frame) as DV25. DV50 uses 4:2:2 subsampling; a CM contains just two luma blocks instead of four. Space that in DV25 would have been allocated to AC terms of the other two luma blocks is available for overflow AC terms. The corresponding DC terms, mode bits, and class bits are reserved. In DV50, the first four bits of DV25’s AC coefficient spaces are filled with EOB symbols.

DV100 doubles the DV50 data rate to 100 Mb/s, and accommodates HD image formats and 4:2:2 chroma subsampling: 1280× 1080 downsampled from 1080i or 1080p, or 960× 720 downsampled from 720p.