HTVQEncode

Full Library / Algorithmic / Image Processing / Basic / HTVQEncode
Block Name: HTVQEncode

public class HTVQEncode
extends Transformer

This actor encodes a matrix using Hierarchical Table-Lookup Vector Quantization. The matrix must be of dimensions that are amenable to this method. (i.e. 2x1, 2x2, 4x2, 4x4, etc.) Instead of performing a full-search vector quantization during execution, all the optimal encoding vectors are calculated before hand and stored in a lookup table. (This is known as Table-lookup Vector Quantization). However, for large vector sizes the lookup tables are unmanageably large. This actor approximates a full search VQ by storing the lookup tables hierarchically. The encoding is broken up into stages, and at each stage a number of 2x1 table lookup VQs are performed. For example, starting with a 4x2 vector in the first stage, codebook 0 (which operates on raw pixels) is used 4 times, resulting in a 2x2 vector of codewords. In the second stage, codebook 1 is used twice, resulting in a 2x1 vector. Lastly, a single 2x1 VQ using codebook 2 (which operates on codewords representing 2x2 vectors) returns a single codeword for the 4x2 vector.

The input is an IntMatrixToken corresponding to the block to be encoded. The values in this matrix are assumed to be between 0 and 255. The output is an IntToken with value between 0 and 255. Integers are used here because of the minimal byte support in VisualSim or JAVA. The size of the input matrix should be the same as the parameters blockHeight and blockWidth.

The codebook is specified as a binary file that will be read during initialization. This file actually contains five sets of codebooks and lookups tables. The first set is for 2x1 blocks, the second is for 2x2 blocks, etc. (Thus the supplied codebook is only sufficient for block sizes up to 8x4 pixels.) In each set, the codebook precedes the lookup-tables. The codebook consists of all 256 codevectors, row scanned from top to bottom. The lookup table consists of 64K entries (one for each pair of codewords from the previous stage). Each entry in the lookup table is an 8-bit codeword.

Stage 0: 2x1 block size
codebook = 256 blocks x 2 bytes = 512 bytes
lookup tables = 65536 entries x 1 byte = 65536 bytes
Stage 1: 2x2 block size
codebook = 256 blocks x 4 bytes = 1024 bytes
lookup tables = 65536 entries x 1 byte = 65536 bytes
Stage 2: 4x2 block size
codebook = 256 blocks x 8 bytes = 2048 bytes
lookup tables = 65536 entries x 1 byte = 65536 bytes
Stage 3: 4x4 block size
codebook = 256 blocks x 16 bytes = 4096 bytes
lookup tables = 65536 entries x 1 byte = 65536 bytes
Stage 4: 8x4 block size
codebook = 256 blocks x 32 bytes = 8192 bytes
lookup tables = 65536 entries x 1 byte = 65536 bytes

The supplied codebook was trained using images from the USC image archive and is suitable for most general applications.
For more information here are some interesting references:
A. Gersho and R. M. Gray, Vector Quantization and Signal Compression. Kluwer Academic Publishers, Boston, 1992.
P. C. Chang, J. May, R. M. Gray, "Hierarchical Vector Quantizers with Table Lookup Encoders," International Conference on Acoustics Speech and Signal Processing, pp. 1452-1455, 1985.
M. Vishwanath and P. Chou, "An Efficient Algorithm for Hierarchical Compression of Video," International Conference on Image Processing, vol. 3, pp. 275-279, Nov. 1994.

Since:
VisualSim 0.2

Field Detail

blockCount

public Parameter blockCount

The number of blocks to be encoded during each firing. The default value is one, which will always work, but using a higher number (such as the number of blocks in a frame) will speed things up.

blockHeight

public Parameter blockHeight

The width, in pixels, of the block to encode.

blockWidth

public Parameter blockWidth

The width, in pixels, of the block to encode.

codeBook

public Parameter codeBook

A Parameter of type String, giving the location of the codebook data file relative to the root classpath.