next up previous contents index
Next: wlbg_adap Up: Reference Previous: Fwvq   Contents   Index

fwvq

$ \bigcirc$Name


fwvq Compress an image with vector quantization of wavelet transform




$ \bigcirc$Command Synopsis


fwvq [-r NLevel] [-e EdgeIR] [-b ImpulseResponse2] [-n FilterNorm] [-w WeightFac] [-s ScalQuant] [-u UnifQuantStep] [-m MultiCB] [-x CodeBook2] [-y CodeBook3] [-A ResCodeBook1] [-B ResCodeBook2] [-C ResCodeBook3] [-D ResResCodeBook1] [-E ResResCodeBook2] [-d] [-R TargetRate] [-o Compress] Image CodeBook1 ImpulseResponse [Qimage ]



-r NLevel : Number of level for wavelet tranform

-e EdgeIR : Impulse reponses of edge and preconditionning filters for orthogonal transform (fimage)

-b ImpulseResponse2 : Impulse response of filter 2 for biorthogonal transform (fsignal)

-n FilterNorm : Normalization mode for filter bank

-w WeightFac : Scaling factor for wavelet coefficients

-s ScalQuant : Use uniform scalar quantization for the resume at level NLevelWav if ScalQuant = NLevelWav or all subimages at levels larger than SalQuant otherwise

-u UnifQuantStep : Use UnifQuantStep levels to uniformly scalar quantize the resume or subimages

-m MultiCB : 1: approximative memory allocation procedure, 2: exhaustive memory allocation procedure (default)

-x CodeBook2 : Sequence of codebooks for second class vectors (fimage)

-y CodeBook3 : Sequence of codebooks for third class vectors (fimage)

-A ResCodeBook1 : Sequence of codebooks for residu quantization after quantization with CodeBook1 (fimage)

-B ResCodeBook2 : Sequence of codebooks for residu quantization after quantization with CodeBook2 (fimage)

-C ResCodeBook3 : Sequence of codebooks for residu quantization after quantization with CodeBook3 (fimage)

-D ResResCodeBook1 : Sequence of codebooks for residu quantization after quantization with CodeBook1 and ResCodeBook1 (fimage)

-E ResResCodeBook2 : Sequence of codebooks for residu quantization after quantization with CodeBook2 and ResCodeBook2 (fimage)

-d : Computes distorsion-rate function

-R TargetRate : Target Rate

-o Compress : Compressed Image (cimage)

Image : Input image (fimage)

CodeBook1 : First sequence of codebooks (fimage)

ImpulseResponse : Impulse response of inner filters (fsignal)

Qimage : Quantized image (fimage)




$ \bigcirc$Function Summary


void fwvq (NumRec , Edge_Ri , Ri2 , FilterNorm , WeightFac , NumRecScal , NStep , MultiCB , CodeBook2 , CodeBook3 , ResCodeBook1 , ResCodeBook2 , ResCodeBook3 , ResResCodeBook1 , ResResCodeBook2 , DistRate , TargRate , Output , Image , CodeBook1 , Ri , QImage )

int *NumRec ;

Fimage Edge_Ri ;

Fsignal Ri2 ;

int *FilterNorm ;

float *WeightFac ;

int *NumRecScal ;

int *NStep ;

int *MultiCB ;

Fimage CodeBook2 , CodeBook3 ;

Fimage ResCodeBook1 , ResCodeBook2 , ResCodeBook3 ;

Fimage ResResCodeBook1 ;

Fimage ResResCodeBook2 ;

int *DistRate ;

float *TargRate ;

Cimage Output ;

Fimage Image ;

Fimage CodeBook1 ;

Fsignal Ri ;

Fimage QImage ;




$ \bigcirc$Description


This module compresses a 8 bits graylevel image using a vector quantization algorithm applied to the orthogonal/biorthogonal wavelet coefficients.

A wavelet transform is first applied to the image. If the -b option is not selected, then an orthogonal transform is performed using the filter contained in ImpulseResponse, and (if selected) the special filters for edge processing contained in EdgeIR (only for Daubechies wavelets). If the -b option is selected, then a biorthogonal transform is applied using the filter pair contained in ImpulseResponse and ImpulseResponse2. ImpulseResponse, ImpulseResponse2 and EdgeIR are file with dedicated format (see WCP/data/filter directory).

The -n option enables to control the filter normalisation. FilterNorm must be an integer ranging from 0 to 2 (default is 2 in the orthogonal case, and 1 in the biorthogonal one).

The -r option controls the number of level of wavelet transform (WavLev must be a positive integer). If not activated, then the number of levels is taken to be equal to the number of levels in CodeBook1.

The -w option enables to multiply the wavelet coefficients by a different factor WeightFacJ (WeightFac must be a positive floating point number) according to the scale J. This sometimes permits to obtain better psychovisual quality for the reconstructed image.

Once the wavelet transform has been performed, the vector quantization algorithm is applied to wavelet coefficients. Each sub-image is quantized separately using a classified and/or multistaged/residual vector quantization algorithm (see fvq module documentation). The codebooks for the first class are contained in the CodeBook1 file. The codebooks for the second and third classes are contained respectiveley in the CodeBook2 and CodeBook3 files. ResCodeBook1, ResCodeBook2 and ResCodeBook3 contain codebooks for the quantization of the residual vectors coming from the quantization with CodeBook1, CodeBook2 and CodeBook3 respectively. ResResCodeBook1 and ResResCodeBook2 contain codebooks for the quantization of the residual vectors coming from the quantization with ResCodeBook1 and ResCodeBook2 respectively. Compress is the output compressed file. Qimage is the quantized image, which can be reconstructed from Compress.

It is possible to apply uniform scalar quantization instead of vector quantization to the upper level subimages with the help of the -s option (see fscalq module documentation for further details). If ScalQuant is equal to the number of level in the wavelet transform, then only the average subimage is scalar quantized. Otherwise all subimages at level strictly greater than ScalQuant are scalar quantized. With the -u option, it is possible to specify the number of steps for the quantization of the average subimage.

The -m option enables to select what will be the input to the memory allocation algorithm between the different subimages. It only makes sense when the codebook files contain more than one codebook per sub-image.

The input to the memory allocation algorithm is a set of discrete rate distortion curves, one for each sub-image in the wavelet transform and each sequence of codebooks (note that in CodeBook1, there is a sequence of codebooks of different sizes for each sub-image, and the same holds for CodeBook2, CodeBook3, ResCodeBook1, a.s.o.). Each point in a rate distortion curve corresponds to the rate and m.s.e. obtained when quantizing a given sub-image with one codebook picked up in a given file CodeBook1, CodeBook2, CodeBook3, ResCodeBook1, a.s.o.. The goal of the algorithm is to select one codebook in each sequence (note that there may be several sequences of codebooks for one sub-image : one in CodeBook1, one in CodeBook2, a.s.o.) in order to minimize the resulting total m.s.e. with the constraint that the total rate remains smaller than a given target bit rate. It begins by taking the smallest codebook in each sequence (typically, it is a size one codebook). The global rate is then 0.0 (if we omit the header inserted at the beginning of the output compressed file, whose size is negligible except at extremely low bit rates). The algorithm works in a greedy fashion, replacing at each step one codebook by a larger one picked up in the same sequence. This codebook is chosen in order to get the best improvement in terms of rate distortion while respecting the bit-rate constraint. If codebook A is replaced by codebook B, then the rate distortion improvement is measured by

$\displaystyle {\frac{{mse_A - mse_B}}{{rate_B - rate_A}}}$.

One thus choose the sub-image, sequence and codebook which maximize this quantity. The algorithm stops when it is not possible anymore to perform a replacement while satisfying the constraint. This algorithm is not optimal, but it is fast and it gives fairly good results.

MultiCB is an integer ranging from 1 to 2. If MultiCB is equal to 2, then exact rate-distortion curves are given as input to the memory allocation algorithm. The problem is then that the running time may be very long, especially if large codebooks are used.

If MultiCB is equal to 1, then approximative rate-distortion curves are given as input to the memory allocation algorithm. It substantially reduces the computation time, but it gives suboptimal allocation. The approximative curves are derived from the rate distortion theory (see [GG92]). They can be written as

mse = a2-2R/N

where R is the rate, N is the dimension of vectors, and a is a constant which depends on the variance of vectors in the corresponding subimage.

The -d option enables to compute a discrete rate-distortion curve.

The -R option specifies the target bit rate for the compression (cancelled if the -d option is selected). TargetRate must be a positive floating point number.




$ \bigcirc$See Also


biowave2, fmse, fscalq, fvq, ibiowave2, iowave2, owave2.




$ \bigcirc$Version 2.02


Last Modification date : Thu Nov 29 20:23:57 2001


$ \bigcirc$Author


Jean-Pierre D'Ales






next up previous contents index
Next: wlbg_adap Up: Reference Previous: Fwvq   Contents   Index
mw 2004-05-05