Wavelet transform is efficient tool for image compression,
Wavelet transform gives multiresolution image
decomposition. which can .be exploited through vector
quantization to achieve high compression ratio. For vector
quantization of wavelet coefficients vectors are formed by
either coefficients at same level, different location or
different level, same location.
This paper compares the two methods and shows that
because of wavelet properties, vector quantization can still
improve compression results by coding only important
vectors for reconstruction. Thus giving priority to the
important vectors higher compression can be achieved at
better quality. The algorithm is also useful for ...view middle of the document...
Size of the vector is variable
and depends on the level of decomposition. It is smaller at
higher level of decomposition. In 141 vector is formed by
Authorized licensed use limited to: Cummins College of Engineering for Women. Downloaded on July 29,2010 at 17:20:02 UTC from IEEE Xplore. Restrictions apply.
Image and Video Coding 1905
PSNR I C
band technique. Cross band technique takes the advantages
of nterband dependency and improves compression.
If we take into consideration HVS response, all the
coefficients are not important for image representation. This
visual redundancy can be removed to improve the
compression ratio further . Edges in the image are more
imponant for good quality image reconstruction. Vectors
giving edge information are more important .Giving priority
to such important vectors embedded coding can be achieved.
PSNR I C
' . PRIORITY BASED ENCODING
Wavelet decomposition represents edges in horizontal
vertical and diagonal direction. If we code only the
coefficients representing edges, image reconstruction at
reduced rate is possible. To find edge region, variance of the
adjacent coefficients can be considered. In vector
quantization if the vectors are formed with adjacent
coefficients from the same band at same
Location, variance of the vectors represent edge region.
Quality of the reconstructed image by
coding only high variance vectors is much better than
interband vector quantiation. Codebook is generated
including high variance vectors from training images. This
results into close match for important vectors and improves
This simple method gives good results even for full search
Variance criterion cannot be used for interband system
because coefficient values show large variations and hence
variance cannot truly identify the true important vectors.
. Results are tabulated in table1
Image was decomposed up to 3 levels using 'Haar wavelet'.
Most coarse version of the image (LL3) IS scalar quantized.
&!!mil Results are...