A. Averbuch et al., IMAGE COMPRESSION USING WAVELET TRANSFORM AND MULTIRESOLUTION DECOMPOSITION, IEEE transactions on image processing, 5(1), 1996, pp. 4-15
Schemes for image compression of black-and-white images based on the w
avelet transform are presented, The multiresolution nature of the disc
rete wavelet transform is proven as a powerful tool to represent image
s decomposed along the vertical and horizontal directions using the py
ramidal multiresolution scheme, The wavelet transform decomposes the i
mage into a set of subimages called shapes with different resolutions
corresponding to different frequency bands, Hence, different allocatio
ns are tested, assuming that details at high resolution and diagonal d
irections are less visible to the human eye, The resulted coefficients
are vector quantized (VQ) using the LGB algorithm, By using an error
correction method that approximates the reconstructed coefficients qua
ntization error, we minimize distortion for a given compression rate a
t low computational cost, Several compression techniques are tested, I
n the first experiment, several 512 x 512 images are trained together
and common table codes created, Using these tables, the training seque
nce black-and-white images achieve a compression ratio of 60-65 and a
PSNR of 30-33, To investigate the compression on images not part of th
e training set, many 480 x 480 images of uncalibrated faces are traine
d together and yield global tables code, Images of faces outside the t
raining set are compressed and reconstructed using the resulting table
s, The compression ratio is 40; PSNR's are 30-36, Images from the trai
ning set have similar compression values compression and quality, Fina
lly, another compression method based on the and vector bit allocation
is examined, The idea is based on allocating different numbers of bit
s to the vectors, depending on their values, encoding the ''type'' of
each vector (large or small) on a bit map, The vectors in each shape a
re grouped and trained together according to the magnitude of their va
riances, A vector that has higher variance is quantized using longer t
ables, Hence, in each shape, the vector coefficients are quantized usi
ng several tables: each vector by the appropriate table, The relation
of each vector to its quantization table is saved in a map file, The m
ajor improvement is achieved by making the process more efficient and
fast since smaller tables are used, and fewer comparisons for locating
the closest vector in the table have to be made, The bottleneck of se
arching large tables, which is very inefficient in all VQ's, is elimin
ated, The compression ratio and the quality of the reconstructed faces
outside the training set have similar results as the previous compres
sion method-compression ratio of 35-36 and PSNR of 35-37-although face
s reconstructed from the training set are slightly better, Distinct wa
velet filters are tested, and the best results are achieved by applyin
g the biorthogonal filters, The results presented here are comparable
with the best results published recently in terms of PSNR.