Image Compression Using Address-Vector Quantization NASSER M. NASRABADI, and YUSHU FENG

1. Image Compression Using Address-Vector QuantizationNASSER M. NASRABADI, and YUSHU FENG IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 38, NO. 12, DECEMBER 1990 Presented by 蔡進義 P9218219

2. Introduction • Vector quantization techniques have been used for a number of years for coding of digital image. • LBG algorithm • The LBG algorithm is very much dependent on the content of the codevectors in the initial codebook and it is local minimum. • Siumlated Annealing (SA) • Address-Vector Quantization • Dynamic A-VQ • Multilayered A-VQ

3. Address-Vector Quantization • Exploit the interblock correlation of the statistical redundancy between the blocks in order to reduce the bit rate. • Address-Vector Quantization • Each codevector represents a combination of address. • Each element of this codevector is an address of an entry in the LBG-codebook. LBG-codebook image A-VQ

4. Address-Vector Quantization • The A-VQ coding system consists of two major components: • A codebook made up of two parts • LBG-codebook • Address-codebook • Four block-transition probability (frequency) matrices each giving the frequency occurrence of two neighboring blocks in • Vertical • Horizontal • 450-diagonal • 1350-diagonal

5. Address-Vector Quantization • The address-codebook is assumed to include all the possible address combination that are encountered during the training process. • The structural information in the image is exploited by the address-codebook to encode four neighboring blocks together as unit. • Only the active region of the address-codebook is addressable by the encoder and decoder. • The most possible address combination

6. Address-Codebook and Block-Transition Probability Matrix

7. Address-Codebook Design • The address-codebook is obtained by dividing all the images in the training sequence into small blocks. • Extract all the possible address combination of four neighboring blocks occurring together in the training sequence. • If the LBG-codebook size is N=128, and the dimension of the codevector in the address-codebook is d=4, then the total possible combination is Nd=1284.

8. Encoding-Decoding Process • The transmitter and receiver have • The same codebook • The same block-transition probability matrices • A score function

9. Encoding • The four neighboring blocks are coded either by the address codebook or by LBG-codebook. • The four neighboring blocks 1, 2, 3, and 4 are first coded by the LBG-codebook to find corresponding address-codevector. • Score parameter P(1/A) x P(2/A) x P(1/B) x P(2/B) x P(1/C) x P(1/D) x P(3/D) x P(1/E) x P(3/E) X P(2/F) x P(4/1) x P(4/2) x P(4/3)

10. Decoding • A simple lookup table consisting of an LBG-codebook and an address-codebook exactly the same as the encoder. • The address-codebook at the transmitter and the receiver have to be in synchronization.

11. Experimental Results Standard VQ Bit rate: 0.437 bits/pixel

12. Conclusion • A new coding technique, address-vector quantization where interblock correlation is exploited. • A score function is used to calculate a parameter to reorder the contents of the address-codebook to bring the most probably address-codevectors into the region of the codebook. • Disadvantages • Synchronization problem • Computational complexity of reordering the contents of the address-codebook during encoding