Standard and Subjective
Images Quality Control of Subband Coded Images using Fuzzy Logic and
Neural Networks
Instructor
Peter Planinsiè
University of Maribor
Faculty of Electrical Engineering and Computer Science
Smetanova 17, 2000 Maribor, Slovenia
E-mail:peter.planinsic@uni-mb.si
Tutorial Presentation
In recent years subband image coding, especially wavelet-subband
image coding, emerged as an efficient method for image compression.
There is also an increasing interest in image compression which provides
desired (and/or optimal) quality (distortion) or bit rate for the
particular application. This requires adequate supervision of the
compression process.
The main aim of this tutioral is to summarise the problems associated
with subband coded images quality measure and control. The past and
our attempts to solve this problems will be presented. The emphasis
will be on our methods, that use fuzzy logic and neural networks.
In recent years subband image coding, especially wavelet-subband
image coding, emerged as an efficient method for image compression
[1, 2, 3, 4]. There is also an increasing interest in image compression
which provides desired (and/or optimal) quality (distortion) or bit
rate for the particular application. This requires adequate supervision
of the compression process.
A special problem in image compression is the criteria for coded
(compressed) image quality that would provide the same subjective
impression. It is well known that at higher compression rates standard
image quality measures do not comply reasonable with the subjective
human eye perception of the image quality.
The main aim of this tutioral is to summarise the problems associated
with subband coded images quality measure and control. The past and
our attempts to solve this problems will be presented. The emphasis
will be on our methods, that use fuzzy logic and neural networks.
The problem in lossy image compression is to find minimum distortion
for constrained bit rate, i.e. compression rate. There can be some
advantages achieved, when weighted distortion is minimised, and where
weights are chosen according to the particular application. In subband
image coding at given filter banks and entropy coding method the optimisation
problem can be formulated as the problem of optimal quantization steps
allocation to the subbands. Assuming high resolution uniform scalar
quantization the bit allocation problem can be solved analytically,
using Lagrange multiplier method or constant slope method, based on
rate-distortion functions of subbands. If the above assumption is
not fulfilled, the analytical solution is no more possible or is not
optimal anymore [5, 6]. In praxis one can still use similar ideas
on operational rate-distortion curves; complex numerical optimisation
methods were proposed [7, 8]. Some of this methods will guarantee
an optimal solution, but are computationally expensive.
To avoid the indicated constrains, we developed an computationally
efficient method, by which the distortion or bit rate is supervised
(or optimised) by using a fuzzy controller and fixed quantization
scale vector. The fuzzy controller is used for adjusting the quantization
steps. Since the coding/decoding process is non-linear, the use of
fuzzy controller allows simple solving of the stability and convergence
problem. Convergence can be additional accelerated with appropriate
selection of initial quantization steps allocation. This can be determined
computationally efficient analytically using the so called low distortion
models, although at higher compression rate the better result is achieved
by fuzzy logic [9, 10, 11, 12, 13].
The most frequently used standard quality measures of compressed
images are mean absolute error (MAE) and quadratic distortion criteria,
such as: mean square error (MSE), peak signal to noise ratio (PSNR)
or signal to noise ratio (SNR). Since at higher compression rates
standard image quality measures do not comply reasonable with the
subjective human eye perception of the image quality, many attempts
have been made to find an appropriate subjective image quality criteria.
S. A. Karunasekara et. al. [14] introduced a composed quality measure
incorporating different types of artefacts and weighted them according
to the sensitivity of the eye. Similar a three-component perceptual
image model for image compression applications was presented [15],
[16]. The use of human visual system model was reported in [17] and
an overview of different quality measures is given in [18]. The multidimensional
quality measures were extensively studied in the past[19]. Additionally,
proposals for the use of fuzzy logic and neural networks were presented
in [20, 21, 22].
To take into account human visual system (HVS), one can use perceptual
bit allocation, i. e. quantization allocation, where the quantization
steps or scale factors are allocated according to the sensitivity
of human visual system (HVS) [23, 24]. This is a complex non-linear
problem. One possibility again is to use fuzzy logic for quantization
scale factor allocation. In [22] this was done for DCT-transform image
coding. Similar approach we used for wavelet-subband image coding
[25]. The quantization scale factors, i. e. scale vector, were obtained
by fuzzy rules, exploiting the known results from psyhovisual experiments.
There fuzzy rules serve as the tool for construction the subband distortion
profile.
In order to introduce closed loop control one needs an adequate quality
measure of compressed images. As a measure of image quality that is
used in the control loop we use standard quadratic quality measure.
Additionally we introduce a novel quality measure that is based on
rate-distortion function, subjective criteria and obtained with fuzzy
logic [26]
Alternatively neural networks can be used for bit or quantiaztion
steps allocation. In our approach neural network was trained to solve
this non-linear problem. There the subband quantization steps were
directly obtained by neural network from the statistical properties
of image and the subbands and desired image quality [27, 28].
Author Presentation
Peter Planinsiè received B. S , M. S Ph. D degree from
University of Maribor, Slovenia in 1979, 1991, and 2000 respectively.
From 1981 to 1986, he was a Staff Scientists at Elektrokovina Maribor
co. Since 1986, he has been working at the University of Maribor,
where he is a senior lecturer for Electrical and Computer Engineering.
In September 1996 he was a visit researcher at German Aerospace Agency
DLR/DFD, Oberphafenhoffen for preparing collaborative project on SAR-images
compression. In August 1998 he was a visit lecturer at Slovak University
of Technology, Bratislava (CEEPUS mobility). He was a program chair
of International conference IWSSIP2000 in Maribor. He had an invited
lecture on International Summer School, that was held at Slovak University
of Technology, Bratislava in August 1997 and invited paper on joint
WorldSES International Conferences NNA2001, FSFS 2001, EC 2001 that
were held in Puerto de la Cruz, Canary Islands in February, 2001.
His research interests include digital signal processing, image compression,
and fuzzy-neural networks applied to the image quality. He is an author
and co-author of many scientific papers and research projects. He
is a member of SPIE and IEEE.
|
