MPEG (Moving Pictures Expert Group)

• 1988 working group within ISO
• borrows significantly from JPEG (lossy)
• encoder is algorithmic and adaptive
• decoder is "dumb"
• perfect for broadcasting (few encoders vs. many decoders)
• MPEG defines the bitstream - not encoders/decoders

Pathway to compression

• spatial and temporal redundancy
  • pixel values are not independent
    • correlated to neighbors in the same frame
    • correlated to neighbors across frames
• psychovisual redundancy
  • human eye has a limited response to fine spatial detail
  • human eye less senstive to detail around object edges

Spatial redundancy

• transform to the frequency domain
  • Fourier analysis - any periodic waveform can be reproduced by adding together an arbitrary number of harmonically related sinusoids of various amplitudes and phases
    • does not result in compression (usually increases)
    • samples are not periodic

STFT (short-time Fourier Transform)

• break up the continuous time domain with windows
  • rectangular
  • # of frequencies depends on size of window
• wrap samples into ring to make "continuous"
• DFT (discrete Fourier transform)
  • we want # input samples == # of frequency coefficients
  • FFT is a fast way to compute DFT

DCT (Discrete Cosine Transformation)

• special case of DFT where sine components eliminated
• repeat original samples in time-reversed order
• perform DFT (8x8 block of pixels)
• produces as many usefull coefficients as input samples

Compression?

• 64 pixels -> 64 coefficients
• but
  • most coefficients will be 0
  • statistically, the further from the top-left, the smaller the magnitude
• compression
  • run-length coding
  • Huffman coding gain
  • quantization of coefficient wordlengths
    • amount weighted according to visibility by a human observer
  • not reversable in the decoder (lossy)

Sequences

• try to predict the next picture from a previous picture
• Send a difference picture
  • difference between coded picture and next picture
    • original picture not available at the decoder
• should also contain spatial redundancy
  • encode the difference picture before sending

Motion-compensated inter-frame prediction

• divide screen into areas called macroblocks (16x16)
• each macroblock steered by a motion vector
• vector gives offset in another picture to find pixel data
  • fetch from another picture
• motion vector overhead can account for 1/3 of bitrate in a "high-action" sequence
• practical search ranges between +/-15 and +/-32 pixels

Types of pictures

• I or 'intra' pictures
  • coded w/o reference to any other pictures
  • just reduce spatial redundancy
  • access points in the bitstream where decoding can begin
• P or 'predictive' pictures
  • use previous I or P for motion compensation
  • each block is either predicted or intra-coded
• B or 'bidirectionally-predictive' pictures
  • can use the previous and/or next I/P for motion compensation
  • can cause a reorder from natural display order
  • prediction is a fetch operation

Sequences of pictures

• a sequence may consist of almost any pattern of I, P, and B pictures
• in a typical sequence, relative sizes of pictures are
  • I = 3*P
  • P = 1.5*B

Summary

• lossless compression generally limited to around 2:1
• compression "sweet spot" of MPEG is between 8:1 and 30:1
• standard VHS - approx 1.5 Mbit/sec
• broadcast NTSC - approx 3 Mbit/sec
• sports/high temporal activity - approx 5-6 Mbit/sec
• betacam (90 Mbit/sec) - approx 10 Mbit/sec

DBS

• broadcast multiple channels over a transponder
  • 23 Mbit/sec available
  • mix/match to balance entropy
  • 1997 - 5-6 channels per trans.
  • 1999 - 8-10 channels per trans.
  • 2001 - 11+ channels per trans.
• take advantage of variable bit rates
  • statistical multiplexing
  • probability that all channels reaching peak entropy at once is very small

DVDs

• 12cm dia disc, short laser wavelength, finer track pitch, better optics
  • approx. 5GB of storage -> movie at 5Mbit/sec
  • disc reduced to 0.6mm (can glue 2 together for 1.2mm disc)
    • approx 10GB of storage -> movie at 10Mbit/sec
• film performs well with MPEG
  • frame rate of 24Hz vs 30Hz (20% savings)
  • film source progressive (no interlacing for motion estimation)
  • pre-digital source oversampled (very high quality signal)
  • focusing, motion blur, etc from film more amenable to the transform and quantization of MPEG