proposed (Luong et al., 2005). A blocking effect
visibility measure based on the local contrast is used
to control the iterative process. In (Singh et al.,
2007), a new technique based on a frequency
analysis is proposed for detecting blocking effects.
The artefacts are modelled as 2-D step function
between the neighbouring blocks. The presence of
the blocking artefacts is detected by using block
activity signal based on HVS and block statistics.
Several other interesting methods (Castagno et
al., 1998; Lee et al., 1998; Zeng, 1999), dealing with
blocking effects have been also developed.
However, most of these studies aim to detect or
remove the blocking effects on the compressed
images.
Here, we propose a different approach which
allows to predict the visibility of the blocking effects
on images prior to compression. The paper is
organized as follows. Section 2 presents the
motivations and describes in details the weighting
procedure. Section 3 is dedicated to the results and
the performance evaluation of our method. The last
section contains the conclusion and perspectives.
2 MOTIVATION AND METHOD
The continuing development of high resolution
imagery technology leads to higher bit rates because
of the increase in both spatial resolution and
intensity range. Much research on block-based lossy
image compression is still needed. However, lossy
compression at low bit rate may produce some
annoying artefacts limiting thus their efficiency.
Here, we focus the study on blocking effects. One of
the main issues related to image compression is how
to control these artefacts. One way to achieve this
goal is first to predict this structured distortion and
then to propose a solution for reducing it. Inspired
by the fact that human observer is able to detect and
recognize blocking effect, even in the absence of the
original image, we propose a new approach based on
a learning process. The approach use here is based
on a training offline process. The main idea is to
compute a weighting function which assigns to each
pixel a weight that could be interpreted as a
prediction probability of the appearance of the local
blocking effect. The main idea developed here is to
study the relation between the appearance of the
blocking effects and the pixels neighbourhoods in
the non-compressed image. Therefore, we perform a
learning offline process on a database containing
various grey-level and color images. The whole
weighting process is summarized in fig. 1.
2.1 Learning Process
The learning process is applied on a database of 211
different real images (from F. C. Donders Centre for
Cognitive Neuroimaging database). These images
contain various kinds of textures with different
roughness and regions with different intensity
distribution and uniform regions.
Figure 1: Synoptic.
First, we analyze the spatial distribution of the
pixels before extracting some local characteristics
from these images. Indeed, the appearance of a
blocking effect in an image area depends highly on
the local descriptors such as color, homogeneity,
gradient etc. These local characteristics could be
expressed in the transformed domain such as
Wavelet Transform or DCT. To make the method
independent of compression method, we use the
local variance as a local homogeneity measure. For
each image f taken from the learning database, we
compute the corresponding local variance image V
(see fig. 2.b). Once the local variance image
computed, we analyse the compression effect in
terms of blocking appearance. To do this, we have to
detect the blocking effects on the compressed
images. Let us define:
•
the compressed images of an original
image f of the database where q represents
the different quality factors (q [1,100] for
JPEG compression).
•
the gradient absolute values images of
.
Depending on the bit rate, the blocking effect tends
to create large uniform zones where the gradient is
null. The blocking effects on the compressed images
could be then detected by analyzing the signal
.
This first simple process gives coarse detection of
blocking effect (see fig. 2.c). We will show that by
Original Image Compressed Image
Local Variance
Gradient image
Image Fusion step
Accumulation matrix computation
Weighting process
SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications
198