UCB Source Coding Jean Walrand EECS Outline UCB

UCB Source Coding Jean Walrand EECS

Outline UCB n n Compression Losless: n n n Audio: n n n Huffman Lempel-Ziv Examples Differential ADPCM SUBBAND CELP Video: n n Discrete Cosine Transform Motion Compensation

UCB n Goal: l n Compression Reduce the number of bits to encode source Approaches: l l Lossless: For data Lossy: For voice, video

UCB n n n Huffman Encoding Lossless Key Idea: Use shorter code words for more frequent symbols EX 1:

UCB n Huffman Encoding EX 2: (continued)

UCB n n n Huffman Encoding (continued) If the symbols are independent and identically distributed, the Huffman encoding is the prefix-free code with the minimum average number of bits. Note: The Shannon encoding requires fewer bits, but requires encoding large blocks of symbols. Both codes assume that the distribution is known.

UCB n n n Lempel-Ziv Lossless Symbols are not independent Distribution is not known Want to minimize the average number of bits Typical application: any file Approach: Build dictionary and replace string with location of prefix in the dictionary

UCB n Lempel-Ziv Example: (continued)

Audio UCB n Examples: n Speech: PCM n ADPCM n SBC n VSELP-CELP n n 64 kbps 32 -64 kbps 16 -32 kbps 2. 4 -8 kbps Audio: PCM n MPEG n 1400 kbps 48 -384 kbps

UCB n Audio (c’d) Differential Encoding (also used for Video): Key Idea is that differences between successive samples may be small n Difficulty: Error Propagation n

UCB n Audio (c’d) Differential Encoding (c’d)

UCB n Audio (c’d) ADPCM: Adaptive Differential PCM Predict next value, encode error

UCB n Audio (c’d) Sub-Band Coding: Improves performance

UCB n Audio (c’d) CELP (Code Excited Linear Predictor)

Video UCB n Discrete Cosine Transform n Objective: Extract “Visible Information” f(x, y) = Sm, n F(m, n) cos(mx) cos(ny)

Video UCB n (cd) Motion Compensation n n Idea: Track motion of picture Encode (motion vector, modification)