IMPROVING MULTISCALE RECURRENT PATTERN IMAGE

CODING WITH DEBLOCKING FILTERING

Nuno M. M. Rodrigues

1,2

, Eduardo A. B. da Silva

, Murilo B. de Carvalho

ergio M. M. de Faria

1,2

, Vitor M. M. da Silva

1,5

Instituto de Telecomunicac¸

oes, Portugal;

ESTG, Instituto Polit

ecnico Leiria, Portugal;

PEE/COPPE/DEL/Poli, Univ. Fed. Rio de Janeiro, Brazil;

TET/CTC, Univ. Fed. Fluminense, Brazil;

DEEC, Universidade de Coimbra, Portugal.

Keywords:

Multidimensional Multiscale Parser, MMP-Intra, Deblocking Filter, Image Coding.

Abstract:

The Multidimensional Multiscale Parser (MMP) algorithm is an image encoder that approximates the image

blocks by using recurrent patterns, from an adaptive dictionary, at different scales. This encoder performs well

for a large range of image data. However, images encoded with MMP suffer from blocking artifacts. This

paper presents the design of a deblocking ﬁlter that improves the performance the MMP.

We present the results of our research, that aims to increase the performance of MMP, particularly for smooth

images, without causing quality losses for other image types, where its performance is already up to 5 dB

better than that of top transform based encoders. For smooth images, the proposed ﬁlter introduces relevant

perceptual quality gains by efﬁciently eliminating the blocking effects, without introducing the usual blurring

artifacts. Besides this, we show that, unlike traditional deblocking algorithms, the proposed method also

improves the objective quality of the decoded image, achieving PSNR gains of up to about 0.3 dB. With

such gains, MMP reaches an almost equivalent performance to that of the state-of-the-art image encoders

(equal to that of JPEG2000 for higher compression ratios), for smooth images, while maintaining its gains

for non-smooth images. In fact, for all image types, the proposed method provides signiﬁcant perceptual

improvements, without sacriﬁcing the PSNR performance.

1 INTRODUCTION

The success of the current state-of-the-art transform-

quantisation based encoders results from their excel-

lent performance in the compression of natural im-

ages. Nevertheless, the relative performance of these

encoders decreases noticeably when we deviate from

the smoothness assumption, as is the case for images

like text, compound (text and graphics), computer

generated, texture, medical, among others. Indeed,

it is a well known fact that most of the encoders that

achieve top results for these image classes have poor

performances for smooth images.

The Multidimensional Multiscale Parser (MMP)

(de Carvalho et al., 2002) is a lossy multidimen-

sional signal encoder, that, unlike most state-of-the-

art image encoders, is not based on the transform-

quantisation paradigm. It is a multiscale recurrent

pattern matching method, that uses an adaptive dictio-

nary for approximating blocks of the original signal.

Using the same pattern matching paradigm, a new

image encoding method, that combines MMP with

the prediction techniques of H.264/AVC (Joint Video

Team (JVT), 2005), was proposed in (Rodrigues

et al., 2005). MMP-Intra is able to achieve quality

gains over the original MMP algorithm for all im-

age types, but particularly for smooth images, where

the performance of MMP is inferior to that of the

top transform-quantisation based encoders. Experi-

mental results show that, when combined with con-

venient dictionary design techniques, the rate distor-

tion (RD) performance of MMP-Intra becomes only

marginally inferior (about 0.2 to 0.5 dB) to that of

the JPEG2000 (Taubman and Marcelin, 2001) and

H.264/AVC high proﬁle (Joint Video Team (JVT),

2005) image encoders, for the coding of smooth im-

ages (Rodrigues et al., 2006). For other types of im-

ages, MMP-Intra consistently maintains its excellent

performance, achieving gains over standardised state-

of-the-art encoders that range from 1 to 5 dB.

MMP-Intra, as MMP, uses the concatenation of the

approximations of the original image blocks, at differ-

ent scales. This process introduces blocking artifacts

in the decoded image, that are particularly evident for

higher compression ratios.

This paper presents a new deblocking scheme for

118

M. M. Rodrigues N., A. B. da Silva E., B. de Carvalho M., M. M. de Faria S. and M. M. da Silva V. (2006).

IMPROVING MULTISCALE RECURRENT PATTERN IMAGE CODING WITH DEBLOCKING FILTERING.

In Proceedings of the International Conference on Signal Processing and Multimedia Applications, pages 118-125

DOI: 10.5220/0001572701180125

 SciTePress

MMP-Intra, that improves the performance of this

image encoder for smooth images, without compro-

mising its compression performance for other image

types. The proposed method is based on a deblocking

method, originally proposed for MMP and a matching

pursuit based multiscale algorithm (de Carvalho et al.,

2002)(de Carvalho et al., 2000), but introduces new

adaptive features, that allow it to optimise the percep-

tual results, as well as the objective performance of

the encoded image.

The experimental results presented in this paper

demonstrate that when the new method is combined

with proper strategies to control the deblocking ﬁlter’s

parameters, it is able to consistently improve the ob-

jective results for smooth images, achieving gains that

go up to about 0.3 dB. For smooth images, these gains

in PSNR correspond to obvious improvements in the

perceptual quality, resulting from the reduction of the

blocking effects introduced by the encoding process.

For non smooth images, like text and compound im-

ages, the new ﬁltering strength control procedure is

able to attenuate, or even eliminate, the smoothing ef-

fects of the deblocking process, that result in a loss of

objective quality.

In the next section we brieﬂy present the MMP and

MMP-Intra image encoding methods. Section 3 de-

scribes a recently proposed dictionary design tech-

nique and explains its importance in increasing the

performance of the MMP-Intra encoder. Section 4

presents the new deblocking strategies proposed in

this paper and is followed by section 5 where the ex-

perimental results of this method are presented . Sec-

tion 6 ends the paper with some closing remarks and

conclusions.

2 IMAGE CODING WITH MMP

A brief discussion of the application of the MMP and

MMP-Intra algorithms to image coding is presented

in this section. More information about these methods

can be found respectively in (de Carvalho et al., 2002)

and (Rodrigues et al., 2005).

2.1 The Mmp Algorithm

MMP is an multiscale approximate pattern matching

algorithm. It approximates an original square im-

age block, or its successive binary segmentations, us-

ing a vector from an adaptive dictionary D. Scale

transformations are used to adapt the dimensions of

blocks with different sizes. The successively seg-

mented blocks, X

, are represented by a binary seg-

mentation tree, where each original square block is

segmented ﬁrst in the vertical, then in the horizontal

direction. The superscript l means that the block X

belongs to scale l or level l of the segmentation tree

(with dimensions (2

⌊

l+1

⌋

× 2

⌊

⌋

)).

A simple deﬁnition of the MMP algorithm can be

given by the following main steps.

For each block of the original image, X

1. ﬁnd the dictionary element S

that minimises the

Lagrangian cost function of the approximation,

given by: J (T ) = D(X

, S

) + λR(S

), where

D(.) is the sum of square differences (SSD) func-

tion and R(.) is the rate needed to encode the ap-

proximation;

2. parse the original block into two blocks, X

l−1

and

l−1

, with half the pixels of the original block;

3. apply the algorithm recursively to X

l−1

and X

l−1

until level 0 is reached;

4. based on the values of the cost functions deter-

mined in the previous steps, decide whether to seg-

ment the original block or not;

5. if the block should not be segmented, use vector S

of the dictionary to approximate X

;

6. else

(a) create a new vector S

new

from the concatenation

of the vectors used to approximate each half of

the original block: X

l−1

and X

l−1

;

(b) use S

new

to approximate S

;

new

to update the dictionary, making it

available to encode future blocks of the image.

i2 i3

Figure 1: Segmentation of a block and corresponding binary

tree: the root corresponds to a original 4×4 block (level 4),

while nodes i

and i

(1×1 blocks) belong to level 0.

This algorithm results in a binary segmentation tree

that represents each original image block. This tree,

represented in ﬁgure 1, is encoded using a top-bottom

preorder approach. In the ﬁnal bit-stream, each leaf

is encoded using a binary symbol ’1’ and followed by

an index, that identiﬁes the vector of the dictionary

that should be used to approximate the corresponding

sub-block. Each tree node is encoded using the binary

symbol ’0’. The string of symbols that represents the

segmentation tree is encoded using an adaptive arith-

metic encoder.

Unlike conventional vector quantisation (VQ) algo-

rithms, MMP uses approximate block matching with

scales and an adaptive dictionary.

IMPROVING MULTISCALE RECURRENT PATTERN IMAGE CODING WITH DEBLOCKING FILTERING

119

Every concatenation of two dictionary blocks of

level l − 1 results in a new block, that corresponds

to a pattern that did not exist in the dictionary and is

used to update it, becoming available to encode future

blocks of the image, independently of their size. This

updating procedure efﬁciently adapts the dictionary,

by using only information that can be inferred by the

decoder, since it is based exclusively in the encoded

segmentation ﬂags and dictionary indexes.

MMP uses a separable scale transformation T

adjust the vectors’ sizes before attempting to match

them, allowing for the matching of vectors of different

dimensions. For example, in order to approximate an

original block X

using one block S

of a different

scale of the dictionary, MMP ﬁrst determines S

[S]. Detailed information about the use of scale

transformations in MMP is presented in (de Carvalho

et al., 2002).

2.2 The Mmp-intra Algorithm

MMP-Intra combines the original MMP algorithm

with predictive coding. For each original block, X

MMP-Intra determines a prediction block, P

, us-

ing previously encoded image pixels and then it de-

termines a residue block, given by R

= X

− P

This residue block is then encoded using MMP.

MMP-Intra uses essentially the same prediction

modes deﬁned by H.264/AVC for Intra coded blocks

(Joint Video Team (JVT), 2005)(Rodrigues et al.,

2005). Intra prediction is also used hierarchically

for blocks of dimensions 16×16 down to 4×4 (cor-

responding to levels 8 to 4 of the segmentation tree).

By the use of the Lagrangian RD cost function, the

encoder jointly optimises the block prediction and the

MMP residue encoding, determining the best trade-

off between the prediction accuracy and the additional

overhead introduced by the prediction data.

MMP-Intra encodes some additional information

for the block prediction, namely the used prediction

mode, m, and the block size used for the prediction

step. This information is used by the decoder to deter-

mine the same prediction block, P

, that was used in

the encoder. This block is added to the decoded resid-

ual block,

, in order to reconstruct the decoded

image block, given by

= P

. Details about

MMP-Intra can be found in (Rodrigues et al., 2005).

3 EFFICIENT DICTIONARY

DESIGN FOR MMP-INTRA

MMP-Intra, as MMP, uses an initial dictionary con-

sisting of a few blocks with constant value. This

highly sparse initial dictionary is very inefﬁcient, but

the updating procedure quickly adapts its blocks to

the original images’ patterns, by introducing new

blocks, S

new

, created by the concatenation of two

vectors of level l − 1 of the dictionary.

Experimental studies have shown that the ﬁnal

number of blocks for each level of the dictionary is,

by far, much larger than the total number of blocks

that are actually used. This difference grows with the

target bit-rate, but can be observed for different im-

age types and target compression ratios. The exag-

gerate growth of the dictionary has the disadvantage

of increasing the dictionary’s indexes’ entropy, com-

promising the method’s performance.

In (Rodrigues et al., 2006), a new algorithm was

proposed to limit the dictionary growth, that intro-

duces a “minimum distance condition” between any

two vectors of each level of the dictionary. This pro-

cess avoids that new vectors, very close to those al-

ready available in the dictionary space, are used to

update the dictionary, by using a new test condition

in the dictionary update procedure. With this new al-

gorithm, a new block of level l, S

new

, is only used to

update the dictionary if its minimum distortion, in re-

lation to the blocks already available in the dictionary,

is not inferior to a given threshold d.

The optimum value for d is a function of the target

bit-rate and therefore of the parameter λ, and must

be carefully chosen. If this value is too small, the

aim of controlling the dictionary growth will not be

achieved, and if it is too large, the dictionary will lose

its efﬁciency in approximating the images’ patterns.

A simple expression for d(λ) (see eq. 1) was de-

termined by the use of a test image set, and allows

the encoder to automatically achieve a close to opti-

mum RD relation, for any given target bit-rate. Fur-

ther details on how this equation was determined can

be found in (Rodrigues et al., 2006).

d(λ) =







5 if λ ≤ 15;

10 if 15 < λ ≤ 50;

20 otherw ise.

(1)

In (Rodrigues et al., 2006), the authors also show

that the dictionary’s indexes can be more efﬁciently

encoded by using a context adaptive arithmetic en-

coder. The dictionary indexes are divided into groups,

according to a context criterion, that, for MMP-Intra,

is the original scale of the block. Instead of using just

one symbol to encode a dictionary index, each index

is transmitted using one context symbol followed by

an index, that chooses among the elements of the cor-

responding segment. This carefully chosen segmen-

tation criterion further explores the statistical depen-

dencies of the MMP symbols, generating gains in the

arithmetic coding module.

SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA

APPLICATIONS

120

4 THE DEBLOCKING FILTER

The MMP-Intra algorithm uses the concatenation of

several approximations of the image blocks, at differ-

ent scales. For each approximation,

, the RD con-

trol algorithm only controls the distortion for the im-

age block and makes no consideration regarding the

continuity in the border of the blocks. This intro-

duces blocking artifacts in the reconstructed image,

that originate from the discontinuities in the block

boundaries.

In this work we present the results of our investi-

gation in deblocking techniques that increase both the

objective as the perceptual quality of the MMP-Intra’s

reconstructed image. We use an adaptive space-

variant ﬁnite impulse response (FIR) ﬁlter to attenuate

the blocks’ borders discontinuities.

Let

X be the reconstructed signal.

X can be re-

garded as the concatenation of several blocks,

that represent the algorithm’s approximation of the

various adjacent areas of the image. The K blocks

, used in the approximation, have no overlapping

areas and different block sizes, given by (2

⌊

⌋

⌊

⌋

). The decoded image can thus be represented

X =

K−1

k=0

(x − x

, y − y

). (2)

In equation 2, each block

corresponds to a dic-

tionary block, of scale l

, that was created by the

MMP dictionary update process. This means that

each of these blocks can be further decomposed into

their basic components, D

, each belonging to the

original dictionary, i.e.

J−1

j=0

. (3)

The blocks D

can be regarded as the basic

”building units” that were used by the MMP-Intra en-

coder and the l

values represent the scale that was

used by the encoder to represent each area of the im-

age. The border points between each of these blocks

correspond to the most probable areas for discontinu-

ities in the decoded image.

In this work we apply a running bi-dimensional

FIR ﬁlter to the reconstructed image,

X. The ﬁl-

ter’s kernel dimensions are successively adapted to

the scale of the original dictionary block that was used

to approximate the area of the image that is currently

being deblocked.

Blocks D

of large scales have larger support re-

gions, meaning that the corresponding area of the im-

age is smoother, while blocks with small values of

are used in more detailed image areas. The used

space-variant ﬁlter has the ability to adapt its support,

and smoothing strength, to the dimensions of each im-

age segment, D

, being considered.

f ( ) f ( )

f ( )

Figure 2: The deblocking process uses an adaptive support

for the FIR of the ﬁlters used in the deblocking.

Figure 2 has a unidimensional representation of a

reconstructed portion of the image, that was approx-

imated by the concatenation of three basic blocks,

), with different scales: l

, l

and l

At each ﬁltered pixel, represented in the ﬁgure by the

arrow, the kernel support of the deblocking ﬁlter is set

according to the scale l

This process is similar to the one proposed in

(de Carvalho et al., 2000), that uses a running aver-

age ﬁlter and sets the kernel support at each point to

+ 1. This ﬁlter is known for its highly smooth-

ing effect, but the support adaptation process controls

its strength according to the detail level of the region

that is being deblocked. This prevents some of the

blurring artifacts that are usually caused by the use

of too powerful deblocking techniques, but the origi-

nal ﬁltering process still results in a reduction in ob-

jective quality for smooth images. Another disadvan-

tage of the original process is that it introduces highly

disturbing blurring artifacts in non smooth images,

resulting in a severe decrease in the ﬁnal values of

PSNR. This fact limits the applicability of this ﬁlter,

because there is no practical way of avoiding the blur-

ring of images that do not need deblocking.

In our work, we have adapted this deblocking pro-

cess to MMP-Intra and developed it. This investi-

gation resulted in a more efﬁcient, highly adaptive,

deblocking ﬁlter, that has some important advantages

over the original method, namely:

• it uses a Gaussian kernel with optimised shape and

support for each image, that adapts the deblocking

strength to the image features resulting in percep-

tual, as well as, objective quality gains;

• the kernel shape optimisation means that the ﬁlter-

ing strength is automatically adjusted and can be

set to an arbitrarily low power. This means that,

for non low-pass images, the new process automat-

ically eliminates the highly annoying blurring ef-

fects and the corresponding PSNR losses;

IMPROVING MULTISCALE RECURRENT PATTERN IMAGE CODING WITH DEBLOCKING FILTERING

121

• the new method considers the dimensions of the

neighbouring blocks as well as those of the block

being ﬁltered, eliminating some artifacts that were

introduced by the original method;

• the proposed algorithm monitors the differences in

the frontiers’ pixels’ intensities, in order to avoid

smoothing steep variations that were present in the

original image and do not correspond to blocking

artifacts.

4.1 Adapting Shape and Support for

the Deblocking Kernel

In our investigation we tested different kernels with

various support regions for the deblocking ﬁlter. Ex-

perimental results showed that the use of Gaussian

kernels, instead of the original rectangular ﬁlter, pro-

duces gains in the PSNR value of the decoded image,

as well as the desired effect of eliminating the block-

ing artifacts.

These tests also demonstrated that the quality of the

deblocked image strongly depends on the dimensions

of the support region of the used ﬁlter. In the original

method, this support is set to l

+ 1. We varied this

value and discovered that it is optimal for the running

average ﬁlter, but that this is not the case when we use

a Gaussian kernel.

Instead of adjusting the support region of the Gaus-

sian kernel, we set the ﬁlter length at the same l

+ 1

samples used in the original method, but adjust the

Gaussian’s variance, producing ﬁlter kernels with dif-

ferent shapes. Consider a Gaussian ﬁlter, with vari-

ance σ

and length L, with an impulse response (IR)

given by:

(n) = e

−

(

n−

L−1

)

2.σ

, (4)

with n = 0, 1, ..., L − 1. We controlled the shape of

the ﬁlter by changing a ﬁlter parameter α, that con-

trols the variance of the Gaussian, by using the ex-

pression:

(n) = e

−

(

n−

L−1

)

2.(α.L)

, (5)

to determine the ﬁlters’ IR.

Figure 3 represents the shape of a 17 tap ﬁlter for

the several values of parameter α represented in the

legend. This ﬁgure clearly demonstrates the explored

relation between the ﬁlters’ shape and their approxi-

mate support. By varying the value of the ﬁlter’s α pa-

rameter, one is able to efﬁciently adjust its IR from an

almost rectangular ﬁlter, with a support region l

+ 1,

to a Gaussian ﬁlter with different lengths. In the limit,

when α tends to zero, the IR of the ﬁlter becomes a

simple impulse, deactivating the deblocking effect for

those cases were it is not beneﬁcial.

The value of the parameter α is controlled by the

MMP-Intra encoder. At the end of the encoding

0.05

0.1

0.15

0.2

0.25

0 2 4 6 8 10 12 14 16

Gaussian FIR of deblocking filters

alpha=0.1

alpha=0.2

alpha=0.3

alpha=0.5

Figure 3: Adaptive FIR of the ﬁlters used in the deblocking.

process, the MMP-Intra encoder tests the deblock-

ing process using different values for the α parame-

ter. It is then able to determine the value that max-

imises the PSNR of the reconstructed image. The

value of α is then appended at the end of the en-

coded bit-stream, by using a 3 bit code, that corre-

sponds respectively to the 8 possible values for α:

{0, 0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.40}. This in-

troduces a marginal additional computational cost in

the encoder, as well as an additional rate overhead,

that is equally negligible.

4.2 Eliminating the Artifacts

Introduced by Deblocking

The original method only considers the dimensions

of the block currently being ﬁltered to set the ﬁlter

support. In our investigation we noticed that this fact

introduces an unexpected artifact, when there exists

the concatenation of wide and short blocks, with very

different intensity values.

This case is represented in ﬁgure 4, where a wide

dark block A is concatenated with two bright blocks:

one narrow block B followed by one wide block C.

When we ﬁlter blocks A and B, a smooth transition

appears, that eliminates the blocking effect in the AB

border. When the block C is ﬁltered, because the used

ﬁlter has a very wide support region, the pixels near

the BC border will suffer from the inﬂuence of some

of the dark pixels of block A. This causes a dark ”val-

ley” to appear in the BC border, that introduces a vis-

ible artifact in the deblocked image.

In order to avoid these artifacts, the new method

controls the ﬁlter length so that the deblocking ﬁlter

never takes in consideration pixels that are not from

the present block or its adjacent neighbours. In the

example of ﬁgure 4, the length of the ﬁlter used in

the C block’s pixels that are near the BC border is

controlled, so that the left most pixel that is used in

the deblocking is always the ﬁrst (left most) pixel of

block B, eliminating the described artifact. In ﬁgure

SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA

APPLICATIONS

122

Block BBlock A

AB BC

Block C

Figure 4: A case were the concatenation of blocks with dif-

ferent supports and pixel intensities causes the appearance

of an image artifact, after the deblocking ﬁltering.

4, this means that the new method uses the ﬁlter rep-

resented by the solid line, instead of the original one,

represented by the dashed line.

Another artifact caused by the original method is

the introduction of smooth transitions in regions of the

image that originally have very steep transitions from

low to high pixel intensity values (or vice versa). The

proposed algorithm monitors the differences in the

frontiers’ pixels’ intensities, in order to avoid ﬁltering

steep variations that do not correspond to blocking ar-

tifacts. This is again controlled by the encoder, using

an adaptive method.

The proposed method uses a step intensity thresh-

old, s, that corresponds to the maximum intensity dif-

ference between the two border pixels, that still allows

for the ﬁltering to occur. This process is represented

in ﬁgure 5, where two blocks A and B with very dif-

ferent intensity values are concatenated. In this case,

the AB border is only ﬁltered if the absolute differ-

ence between the border pixels is inferior to the de-

ﬁned value for s, i.e., |A

− B

| < s.

Block A Block B

Figure 5: A case where a steep variation in pixel intensities

is a feature of the original image.

The value of s is again chosen in order to maximise

the PSNR value for the particular image that is be-

ing deblocked. The encoder tests a set of different

step values and transmits the code corresponding to

the chosen value. A three bit code is again used to

represent the eight possible values for s, belonging to

the set {0, 16, 32, 64, 96, 128, 192, 255}, where s = 0

corresponds to never ﬁltering the borders and s = 255

corresponds to the case where all blocks are ﬁltered.

5 EXPERIMENTAL RESULTS

Experimental tests were performed using the pro-

posed and original deblocking methods. Figure 6

presents a detail of image LENA 512, encoded using

the described MMP-Intra algorithm without using any

deblocking technique and compares it with the results

of using the the deblocking technique from (de Car-

valho et al., 2000), and the new deblocking technique,

proposed in this paper.

Perceptually, we can observe that the new deblock-

ing ﬁlter is able to efﬁciently eliminate the blocking

artifacts in Lena’s face and hat, without compromis-

ing the image quality at regions with high detail, like

Lena’s hair and her hat’s feathers. In this case, the

used ﬁlter has α = 0.10 and s = 255.

When compared with the original deblocking

method, we can observe that the smoothing effect in-

troduced by the proposed method is not as strong,

avoiding the introduction of some blurring artifacts

that are noticeable in the image of ﬁgure 6 b), spe-

cially in the areas with ﬁner details.

In ﬁgure 6 b) we can also observe the ﬁrst type

of artifacts, explained in section 4.2. They appear

in Lena’s shoulder, where the previously described

”dark valleys” are easy to observe. We can see that

the proposed method efﬁciently eliminates these arti-

facts.

Figure 6 b) also shows that the original method, de-

veloped originally for the MMP encoder, suffers from

a unexpectedly high performance loss, when used

with MMP-Intra. Because MMP-Intra uses predic-

tive coding, the used dictionary blocks approximate

residue patterns. In some cases, where the prediction

step is particularly efﬁcient, some detailed areas are

approximated by large, smooth, residue blocks added

with detailed prediction blocks. In this case, the de-

blocking process uses a wide ﬁlter to deblock an im-

age area that is not necessarily smooth. When this

happens, the use of the original deblocking method

originates serious artifacts, like the one observed in

Lena’s lip. Even when this fact is not as obvious as

in the presented case, we can generally say that this

factor seriously compromises the performance of the

original method, when applied to MMP-Intra, result-

ing in a severe reduction in the PSNR results. How-

ever, due to its adaptability, the proposed method does

not seem to suffer from this disturbing factor.

Figure 7 a) shows the objective quality results for

image Lena 512, for the MMP-Intra method with no

deblocking and with the two tested deblocking tech-

niques. Figure 7 b) highlights the PSNR quality gains

introduced by the deblocking ﬁlter, that are more rel-

evant for higher compression ratios, where the block-

ing artifacts are more noticeable. These gains go

up to more than 0.3 dB, allowing for the PSNR re-

sults of MMP-Intra for image Lena to come even

IMPROVING MULTISCALE RECURRENT PATTERN IMAGE CODING WITH DEBLOCKING FILTERING

123

(a) No deblocking (30.92 dB) (b) Original deblocking (28.93 dB) (c) New deblocking (31.21 dB)

Figure 6: A detail of image Lena 512, encoded with MMP-Intra at 0.135 bpp.

closer to the ones of top state-of-the-art transform-

quantisation based encoders, like JPEG2000 (Taub-

man and Marcelin, 2001) and H.264/AVC, (Joint

Video Team (JVT), 2005), shown in ﬁgure 7 a). In

fact, we can see that, for low bit-rates, the proposed

method allows for MMP-Intra to achieve equivalent

results to those of the JPEG2000 algorithm.

We also performed experimental tests using non

smooth images, like text image PP1205 and com-

pound (text and grayscale) image PP1209. Im-

ages PP1205 and PP1209 were scanned, respec-

tively, from pages 1205 and 1209 of the IEEE

Transactions on Image Processing, volume 9, num-

ber 7, July 2000 and are available for download at

http://www.estg.ipleiria.pt/∼nuno/MMP/. These tests

showed that the proposed kernel adaptation algorithm

eliminates the highly disturbing blurring artifacts in-

troduced when the original deblocking techniques are

applied to these images. This can be conﬁrmed in

ﬁgure 8, where the perceptual results for compound

image PP1209 are presented.

Figure 8 c) also shows that the use of the new strate-

gies as a simple post processing deblocking algo-

rithm, allows for a deblocking effect that improves the

subjective quality of the decoded image, at the cost of

a slight reduction in the PSNR value. In addition, it

shows that the second type of artifacts introduced by

the original method, that introduce a smoothing ramp

in areas of the image that originally had an abrupt

variation, is efﬁciently eliminated by the proposed al-

gorithm (in this example, the value of s was set to 32).

Figure 9 shows the PSNR results for compound im-

age PP1209, for the case presented in ﬁgure 8, where

the deblocking process allows for an increased per-

ceptual image quality, at the cost of a small reduc-

tion of the PSNR value. As we can see, even in

this case the objective quality achived by the MMP-

Intra encoder is still about 1 dB better than that of the

H.264/AVC encoder and 2 dB better than that of the

JPEG2000 encoder.

0 0.2 0.4 0.6 0.8 1 1.2 1.4

PSNR (dB)

bpp

MMP-Intra - No Deblocking

MMP-Intra - Deblocking

MMP-Intra - Original Deblocking

H.264/AVC - High

JPEG2000

(a)

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0 0.2 0.4 0.6 0.8 1 1.2 1.4

PSNR gains (dB)

bpp

(b)

Figure 7: a) Objective quality results for image Lena;

b) PSNR gains of the new method, when compared with

MMP-Intra with no deblocking, using α = 0.10, s = 255.

SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA

APPLICATIONS

124

(a) No deblocking (32.23 dB) (b) Original deblocking (27.13 dB) (c) New deblocking (32.04 dB)

Figure 8: A detail of compound image PP1209, encoded with MMP-Intra at 0.61 bpp.

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

PSNR gains (dB)

bpp

MMP-Intra - No Deblocking

MMP-Intra - Deblocking

MMP-Intra - Original Deblocking

H.264/AVC - High

JPEG2000

Figure 9: Objective quality results for image PP1209 (adap-

tive deblocking ﬁlter used α = 0.10 and s = 32).

6 CONCLUSION

In this paper we present a new adaptive deblocking

technique that allows for improvements in both per-

ceptual and objective quality, for the MMP-Intra im-

age encoding algorithm. This method uses a space-

variant FIR ﬁlter with an adaptive shape and support

Gaussian impulse response. The ﬁlter parameters are

automatically controlled in order to maximise the ob-

jective quality for smooth images and eliminate the

disturbing blurring artifacts for non smooth images,

like text and graphics.

The use of the new deblocking techniques achieves

one of the main objectives in the on going research

of multiscale recurrent pattern image encoders: ﬁnd-

ing ways to improve the algorithm’s performance

for smooth images, without compromising its ex-

cellent performance for non low-pass images, like

text and graphics. Experimental results have shown

that, for smooth images, the proposed techniques al-

lows for coding gains that go up to 0.3 dB for low

bit-rates, where the blocking artifacts are more no-

ticeable, achieving the same objective quality as the

JPEG2000 algorithm. Nevertheless, for non low-pass

images, like text and graphics, the proposed method

introduces no losses, allowing MMP-Intra to maintain

its 1 to 5 db advantage over the state-of-the-art image

encoders.

REFERENCES

de Carvalho, M., da Silva, E., and Finamore, W. (2002).

Multidimensional signal compression using multi-

scale recurrent patterns. Elsevier Signal Processing,

(82):1559–1580.

de Carvalho, M., Lima, D. M., da Silva, E., and Finamore,

W. (2000). Universal multi-scale matching pursuits

algorithm with reduced blocking effect. IEEE Inter-

national Conference on Image Processing.

Joint Video Team (JVT), ISO/IEC MPEG & ITU-T VCEG,

I. J. . I.-T. S. Q. (2005). Draft of Version 4

of H.264/AVC (ITU-T Recommendation H.264 and

ISO/IEC 14496-10 (MPEG-4 part 10) Advanced

Video Coding).

Rodrigues, N. M. M., da Silva, E. A. B., de Carvalho, M. B.,

de Faria, S. M. M., and Silva, V. M. M. (2005). Uni-

versal image coding using multiscale recurrent pat-

terns and prediction. IEEE International Conference

on Image Processing.

Rodrigues, N. M. M., da Silva, E. A. B., de Carvalho, M. B.,

de Faria, S. M. M., Silva, V. M. M., and Pinag

e, F.

(2006). Efﬁcient dictionary design for multiscale re-

current patterns image coding. ISCAS 2006 IEEE In-

ternational Symposium on Circuits and Systems.

Taubman, D. S. and Marcelin, M. (2001). JPEG2000: Im-

age Compression Fundamentals, Standards and Prac-

tice. Kluwer Academic Publishers.

IMPROVING MULTISCALE RECURRENT PATTERN IMAGE CODING WITH DEBLOCKING FILTERING

125