Research
My research interests are in image processing, including lossless
and lossy image compression, perceptually lossless image compression,
and texture segmentation. Recently my research focus is on context-based
conditional entropy coding of Vector Quantization(VQ) indexes for
images. This method may open the new way to approach the rate-distortion
optimality for traditional VQ design.
1. Bayesian-based Conditional Entropy Coding of VQ Indexes.
Based on source coding theory, Vector Quantization
can approach the optimal rate-distortion tradeoff as the vector dimension
approaches infinity. However, limited by the inherent computational
intractability of VQ in dimensions, VQ's theoretical promise is yet to
be fully realized in image compression practice because of its modest
vector dimension. This must exist
a large statistical redundancy between the inter blocks. Other than
the general VQ point of view to utilize this property, we propose a new
mothed which directly uses lossless techniques to code the VQ indexes obtaining
from basic VQ. Bayesian rule is
utilized to estimate the conditional probabilities of VQ indexes based on
causal context in order to drive the arithmetic entropy coder, and
the compression performance of this method can compare with the best
VQ performance we have known.
2. Conditional Entropy Coding of VQ Indexes via Context Selection and
Quantization.
Since the block sizes of practical VQ image coders are not large
enough to exploit all high-order statistical dependencies among pixels,
high-order statistical modeling of VQ indexes is presented. Through
various context selection and quantization, one context structure which the
pixels are contiguous to the current block is finally chosen. Based on
such context, the conditional probabilities of VQ indexes can be estimated
so good that the compression gains of this method are 10% higher than
Address VQ(A-VQ), which first
used lossless idea to code VQ index, with the same compression
distortion and a tiny fraction
of A-VQ's computational cost are needed.
This paper is under writing.
3. The Application of Conditional Entropy Coding of VQ Indexes for
High-fidelity Medical Image Compression.
It's well known that VQ is very suitable for such kind of image
compression that the images we want to compress belong to one category,
such as medical image compression. However, low compression ratio impedes
its development. Based on the encouraging results from conditional
entropy coding of VQ index, we tailor this idea to medical images by
refining the context modeling for a given class of medical images. This
image-type-dependent method has obtained compression performance superior
to a state-of-art wavelet image coder on the class of MR brain images.