Author: Chang et al. Title: An Image Coding Scheme Using SMVQ ...

chardfriendlyAI and Robotics

Oct 16, 2013 (3 years and 11 months ago)

82 views

Author: Chang et al.

Title:
An Image Coding Scheme Using SMVQ and Support Vector Machines
.

Manuscript

Number:
NEUCOM
-
D
-
04
-
00084R1


With the reviewer

s
excellent
suggestions, it
is extremely
thankful

and
helpful

to
make
the

manuscript a clearer version
.
We
have
re
-
revised the manuscript according
to reviewer

s
very
valuable comments; t
he

main
revised
portions

are listed as follows:


Reviewer #2:

This paper describes a method for image compression based on the
use of

support vector machines in order to classi
fy each block of an image into

``edge'' or ``non edge''. A codebook is selected in order to compress the

underlying
information.


Comments
:

-

Again the author spends lot of space in reviewing well known concepts

marginal
to the main contribution of the paper
. The paper does not

investigate

theoretically
and experimentally the main contributions in

details.

-

Globally the paper is

clear.




In equation (2) ... the author should mention the variable under the

minimum.



Equation (2) has been modified
according to

the comments
:

[
ORIGINAL
]

.

[MODIFIED]
.




The title ``related theories'' can be rep
laced with a `` review on SVMs...'' or ....

Again the review on SVM, DWT... can be more concise with

less details

and

more
understanding on these materials instead of

restating them...

this is usually what a
reader expects from a journal paper.



As the re
viewer

s suggestion

t
he title of Section 2 is
changed
from the original

Related Theories


into


Review of Background


getting
more appropriate for
the
content
s

of
this section.

However
, d
ue to this paper summarized
different materials

from
several

fields
of
image compression

(side
-
match VQ), image processing

(discrete wavelet transform) and machine learning (support vector machine
s
)
,
we
beli
e
ve
that
the

way

this

section

represent
s

should be

suitable

and
natural
to
understand for
most readers
.




Figure 4 is not necessary.



T
he
figure

titled

s
implifying the classification task by feature mapping


is

removed.




In Figure 5 ... the sentence ``the built corresponding ...'' can be

replaced by `` The
corresponding state codebook''



It
has been

mod
ified

according to

the comments.


Other comments:




title 3.1 `` Feature extracting phase'' can

be replaced by `` Feature

extraction''...



title 3.2 can be replaced just by ``SVM training''



It is
much

appreciat
ed

to
the reviewer’s
valuable
suggestion
s

which
mak
e

this
manuscript
clear
er. The subtitles
have been replaced as well as
other

occurrence

in
the

manuscript
. We simply list the changes as follows:

[ORIGINAL]
As Fig. 5 depicts, we can simply divide the proposed scheme into three
phases: (1) the fe
ature extracting phase, (2) the SVMs training phase, and (3) the
image encoding / decoding phase.

[MODIFIED]

As Fig. 5 depicts, we can simply divide the proposed scheme into three
phases
: (1) feature extracti
on
, (2) SVM training, and (3) the image encoding

/
decoding phase.




In page 4, in the sentence ``a hierarchical three sided
-
side'' there is a

repetition of
the word side.



Because

the title of the
cited paper

is


Three
-
sided side match finite
-
state vector
quantization


[6]
original
ly
,
we

would like t
o
re
tain th
is

sentence

unchanged

for
consistency
.




In the top of page 5, the author mixes active and passive forms in the

same
sentence, it will be better to choose only one of them. In the same

paragraph, the
sentence ``we are going ...'' can be replac
ed just by ``we

will''



The mentioned sentences have been rewritten
according to

the comments. For
your convenience we listed the changes below:

[ORIGINAL]

Then,
in Section 3, we are going to describe the proposed scheme in
detail, and the experimental re
sults will be shown in Section 4.

[MODIFIED] Then, in Section 3, we will describe the proposed scheme in detail. The
experimental results will be shown in Section 4.




The author stated that high frequency coefficients of the wavelet

transform are
redunda
nt ... In fact they are redundant only if their value

is 0.



Yes
, we

completely
agree with the reviewer

s opinions.
Precisely
,

the high

frequency parts can not be treated as redundan
t

but are
imperceptible

to
human eyes.

Thus

w
e rewrote the
relative descr
iption
s

to
avoid

th
is

unnecessary
misunderstanding
:

[ORIGINAL]
on the other hand, to our human eyes the high frequency parts of the
images are not
perceptible

and could be treated as redundant.

[MODIFIED]
on the other hand
,

since the high frequency parts o
f the images are not
very
perceptible

to

human eyes, they may be treated as one kind of redundancy from
this viewpoint.




In the sentence ``of orthonormal wavelets basis...'' the (s) in wavelets

should be
omitted.



We have corrected the
mentioned
sentence.




In page 8, the author should support the works ``on hand
-
written pattern

recognition,
bioinformatics ...'' by some
references.



As the suggestions, additional references
that

support the
se

works
in

fields
of

text
categorization, hand
-
written pattern recognition,
and bioinformatics

are given in
[13
-
15]
:

[13]
I. Guyon, J. Weston, S. Barnhill, and V. Vapnik
,

Gene select
ion for cancer
classification using support vector machines
,


Machine Learning
, vol. 46, 2002,
pp. 389
-
422.

[14]
T
.

Joachims
,
“Text
c
ategorization with
s
u
p
port
v
ector
m
achines:
l
earning with
m
any
r
elevant
f
eatures
,


Proceedings of the 10th European Confere
nce on
Machine Learning
,
Chemnitz, Germany
, April 1998, pp. 137
-
142.

[15]
Y.

LeCun, L.

D. Jackel, L.

Bottou, A.

Brunot, C.

Cortes, J.

S. Denker,
H.

Drucker, I.

Guyon, U.

A. Muller, E.

Sackinger, P.

Simard, and V.

Vapnik
,
“Comparison of
l
earning
a
lgorithms
for
h
andwritten
d
igit
r
ecognition
,


Proceedings of

the I
nternational Conference on Artificial Neural Networks
,
Paris
,
France
,

October
199
5
,

pp. 53
-
60.




In page 8, the sentence ``suppose that we have a given'' can be replaced

just by
``given a training se
t".



The sentence has been rewritten

as the comments to make it clearer
. For the
convenience
,

we listed the changes below:

[ORIGINAL]

as follows: Suppose that we have a given data set of examples
, where

is
an example
instance and
,
.

Here,

is the
-
dimensional input space



[MODIFIED]


as follows: Given a training set
,

i
s an example instance
and
,
.
Here,

is the
-
dimensional input space





In page 9, again the expression `` a kind of benchmark `` is difficult to

understand
and should be clear (for example by adding a footnote).



Since th
is

expression
is

not

very

clear

to understand, we
decide

to
abandon
the
original

version

and
depict

thi
s
concept

with
a
nother

way

which
is

easier to accept
.
However,
it
is

noted that
the

idea
behind

SVM
is to

derive a hyperplane
(OSH)

with
the margin maximized,
and

t
his OSH
is thought

to
be the
one

having the best
generalization abilities
among all possible
hyperplanes theoretically.

[ORIGINAL]


.
and it can be treated as a kind of benchmark for
generalization

abilities of the hyperplane
.

[MODIFIED]

In a sense, it can be realized as the
generalization

abilities of a
hyperplane
: the larger the margin is, the better the generation ability will be.




In page 9,

capital letters in ``Optimal Separating Hyper...'' should be

removed.



T
he capital letters have been removed

as the suggestions
.




In page 11, the sentence

``the SVMs good generalization ..... and

practical'' is very
difficult to understand and technically incorrect...

the generalization is a property not a
tool.



We rewrote the sentence to correct th
e

inappropriate
representation
:

[ORIGINAL]

The SVMs good
generalization ability

will be utilized as the identifier
of edge blocks to make the recognition task more feasible and practical.

[MODIFIED]

B
ecause of its good generalization performance

[16]
, SVM

here
will be
used as the block type (
edge

/

non
-
edge
) ide
ntifier, making the recognition task more
feasible and practical
.




English

might be improved for instance in the upper part of page 12.

T
he sentence
``To gather.... domain'' can be omitted.

In page 12, ``SVM training /classifying stages''
can be replaced

by ``SVM

training and classification''.



T
he sentence of

To gather an image

s energy, it is common in image processing
to transform it from the spatial
domain

into the frequency
domain
.


has been omitted.

As the comments,

in page 12
,

t
he

term

SVM train
ing /classifying stages


was
replaced
by


SVM

training and classification

.

The upper part of page 12 was
re
written

as well
to be

clearer
.





In page 12, why

16 x 16 coefficients make SVM training slow?



It

is

well
-
known that PCA

(
Principal
C
omponent
A
na
lysis
)

method
can be

us
ed

to
reduce the
complexity of problems in
data dimension

and
re
tain

its

maximum

amount
of variation
.

PCA

indeed brings
benefit
s on the aspect of speed

acceleration
.

S
imilarly,
DWT technique

is adopted
here
in the same manner.
Actual
ly
,

in the manuscript
,

we
did not say
that the
original
coefficient

size
make
s

SVM training
slow, but
,

just like
using PCA,
i
f we can
trade a tiny bit of precision loss for a significant

reduction in
complexity in SVM

training
and

classifying computation
,
why not?




In page 13, the sentence after ``t
he applying the transforming''
must be

rewritten.



The

mentioned sentence has been corrected
:

[ORIGINAL] That is, after the applying the
transforming phase discussed in
Subsection 3.1 to each block
,



[MODIFIE
D] That is, after applying the transformation procedure which is discussed
in Subsection 3.1 to each block,






The notion of "edges" / "no edges" is not clear. It is strongly

recommended that
the author shows some images of

what is

``edges'' and

``non
-
ed
ges''.



Some example images of
"edges" / "no
n
-
edges"

are
given

in Figure 4.

Note that
every

block is

rescaled to the
expressible

size
:

[ORIGINAL]

[MODIFIED]


S
ome example blocks of different types are
shown

in Figure
4

for
clarity
.


Fig.
4
:

Example

blocks
with

different typ
es

(a)

edge block
(b)

edge block

(c)

non
-
edge block
(d)

non
-
edge block

(
a
)

(
b
)

(
c
)

(
d
)




In page 19, what is ``GMM'' ?



It’s

the
abbreviation

of the well
-
known classification technique
,

Gaussian Mixed
Model
.

W
e replaced the
abbreviation


GMM


with
its

ful
l name in the manuscript to
avoid

the
confusion

that
possibly

occurs
.



*******
**** END FORWARDED MESSAGE***********