850 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 12, NO.

10, OCTOBER 2002

An Efficient Spatial Prediction-Based Image

Compression Scheme
Chin-Hwa Kuo, Member, IEEE, Tzu-Chuan Chou, and Tay-Shen Wang

Abstract—In this paper, we have designed a spatial predic- hierarchical levels. In the prediction phase, instead of using only
tion-based image-compression scheme. The proposed scheme preceding pixels according to the order of the raster scan as the
consists of two phases, namely, the prediction phase and the quan- input to predict the sample pixels [5], such as lossless JPEG [6],
tization phase. In the prediction phase, a hierarchical structure
among pixels in the image is built. Following the constructed our scheme utilizes the surrounding pixels to predict every cen-
hierarchical structure, the neighboring pixels are utilized to tral pixel. The four cross-neighboring pixels are used in the first
predict every central pixel. The prediction scheme generates an step, and then the four diagonal neighboring pixels are used in
image map which indicates the prediction errors. The structure the second step. The small image composed by the remaining
of the resulting image map is very similar to the result of discrete one-fourth pixels is applied to the above two steps iteratively.
wavelet transform. Thus, most quantization methods of wavelet
or subband image-compression algorithms can be followed in our This resulting image hierarchy is very similar to the hierarchy
scheme directly to yield good compression performance. In the made by DWT in some respects, but they are not entirely equiv-
quantization phase, we design a multilevel threshold scheme to alent. We can then integrate this hierarchy with many quantiza-
further enhance the result of SPIHT by taking the significance of tion methods of wavelet or subband image-compression tech-
the pixel values and the hierarchical levels into account. Further- nologies, e.g., EZW [1] and SPIHT [2] algorithms, to carry out
more, the proposed scheme can be realized only by few integer
additions and bit shifts. Simulation results indicate that the visual the image-compression process in a smooth manner.
quality of the designed efficient spatial prediction-based image We also modify the SPIHT with multilevel thresholds
compression scheme is competitive with JPEG. All the above (which differs from multithreshold multithreshold wavelet
features make the designed image-compression scheme beneficial codec (MTWC) [3]) in the quantization phase. The designed
to the applications of real-time and wireless transmission in multilevel threshold quantization scheme takes the significance
low-computational power environments.
of the pixel values and the hierarchical levels into account in
Index Terms—Image hierarchical structure, predictive image order to further enhance the compression performance. The de-
coding, progressive image compression, quantization with multi- signed image-compression scheme only requires a few integer
level thresholds.
additions and bit shifts in the encoding and decoding processes.
Furthermore, in-place processing is taken, i.e., the prediction
I. INTRODUCTION errors are stored in the same places as the central pixels. Thus,
the temporary storage space required is significantly saved in
I MAGE compression plays an important role in modern net-
worked multimedia applications. Many powerful compres-
sion schemes have been developed in recent years. In particular,
the coding process. Our experimental simulations show that the
performance of the proposed scheme is very competitive with
discrete wavelet transform (DWT) has become the most impor- the baseline JPEG [6], which uses DCT as the core technology.
tant core technology of image compression [1]–[4]. However, As a result, an efficient and progressive compression scheme is
schemes using DWT may suffer because of the computational obtained. These features make the proposed scheme beneficial
complexity. In comparison with the traditional discrete cosine to the applications of real-time and wireless transmission in
transform (DCT) approach, DWT usually needs more computa- low-computational power environments.
tion power. Therefore, this may restrict the applications of DWT The structure of this paper is as follows. Section II presents
in the areas of real-time and wireless multimedia at the present the background of this work. Section III introduces the new spa-
stage. tial prediction approach, and hierarchical structure of an image.
The designed efficient spatial prediction-based image com- The quantization scheme with multilevel thresholds is discussed
pression (ESPIC) scheme consists of two phases: the prediction in Section IV. Finally, experimental results and conclusions are
phase and the quantization phase. Based on the pixel geomet- proposed in Section V and Section VI, respectively.
rical relationship, we separate all pixels of an image into several
II. SPATIAL PREDICTION-BASED IMAGE-COMPRESSION
SCHEME
Manuscript received March, 2000; revised June 21, 2001. This work was sup-
ported in part by NSC of Taiwan under Contract NSC 89-2218-E-032-019. This Generally speaking, two-dimensional (2-D) coding schemes
paper was recommended by Associate Editor H. Watanabe. are based on reducing or eliminating any coding, interpixel, or
C.-H. Kuo and T.-C. Chou are with the Computer and Networking Labora-
tory, Department of Computer Science and Information Engineering, Tamkang psychovisual redundancies. In predictive coding, the difference
University, Taipei, Taiwan, R.O.C. between the original value and predicted value of a pixel is ex-
T.-S. Wang is with the Computer and Networking Laboratory, Department of tracted and coded. The typical spatial coding system is shown in
Computer Science and Information Engineering, Tamkang University, Taipei,
Taiwan, R.O.C. and also with Cyberlink Corporation, Taiwan, R.O.C. Fig. 1. It consists of an encoder and a decoder. Both encoder and
Digital Object Identifier 10.1109/TCSVT.2002.804878 decoder contain an identical predictor. While each successive
1051-8215/02$17.00 © 2002 IEEE
KUO et al.: AN ESPIC 851

Fig. 1. A lossless predictive coding model [5].

pixel of an input image, denoted , is introduced to the encoder,

the predictor generates the corresponding predictive value, de-
noted , from the previous inputs. Then, the prediction error
is calculated as . A symbol encoder such as
Huffman coding, Arithmetic coding, or others, compresses the
prediction errors into a compact data stream. A good predictor
generates prediction errors with a small entropy value reflecting
low energy. Through the symbol encoder, the system can com-
pact the coded image data. At the conclusion the decompressed
image corresponds absolutely with the input image. Fig. 2. Relationship between the sample x and its predictive pixels in lossless
A traditional 2-D spatial predictor is usually represented as JPEG [6].
(1)
By building the hierarchical structure of an image, we break this
limit and extend our scheme to the domain of lossy image com-
pression. Meanwhile, the designed image compression yields
good compression performance and other significant features.
(1) The details are described in Section III.
The prediction scheme processes a raster-scan of the whole
III. PROPOSED SPATIAL-PREDICTION SCHEME
image from left to right and top to bottom. In order to decode
the compressed images, the traditional prediction schemes The proposed image-compression scheme based on spatial
cannot use all pixels around the target pixel to predict the color prediction consists of two phases: the prediction phase and
value of the target pixel. Fig. 2 shows the relationship between the quantization phase. In the prediction phase, the proposed
the neighboring pixels and the current sample pixel. For the scheme employs more information from pixels around every
current sample pixel marked by , only the pixels above the sample pixels depending upon the hierarchical structure among
thick line in Fig. 2 can be used to predict the sample value. pixels in the image. This allows us to convey more information
On the other hand, all pixels below the thick line cannot be from the predictive pixel by using its neighboring pixels. In
utilized for prediction. This is the main disadvantage of the the quantization phase, we modified the SPIHT quantization
image-compression scheme based on spatial prediction, since algorithm to enhance the compression performance. In order to
the predictor cannot use more information from the neighboring construct the hierarchical structure, we first mark every pixel
pixels. Here, we have designed a scheme to overcome this with a number by tiling the whole image with the hierarchical
drawback of the traditional scheme. map denoted in Fig. 3. The following algorithm produces the
The traditional image codecs based on prediction coding are hierarchical map.
usually used in lossless image compression. The coding process Initial:
is through the raster-scan order [5]. That is, preceding pixels are 1. Let and denote height and width of the image. Thus,
used as inputs of the designed predictor to ensure that the values the row numbers are denoted from 0 to ; the column
of all the pixels of the image can be restored in the decoding step. numbers are denoted from 0 to .
852 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 12, NO. 10, OCTOBER 2002

Fig. 3. Hierarchical map (16 2 16 matrix).

(a)

2. Set the variable as 1.

3. Mark all pixels of the whole image as number 0.
Step 1:
Mark all pixels as number if its sum of row number and
column number is even.
Step 2:
Mark all pixels as number if its row number is even and
column number is odd.
Step 3:
If the mapping process is complete, stop computing. Other-
wise, collect all pixels that are still marked as number 0 and
combine them into a new small image. Thus, the size of the new
small image becomes one fourth of the original one. Then re-
assign new row numbers and new column numbers to the new
image. Let become and go back to Step 1.
(b)
In Step 3 of the above algorithm, the stop condition is
heuristic. In our experience, when , Fig. 4. Prediction process of Steps 1 and 2 (8 2 8 matrix).
the stop condition is set up.
After marking the whole image with the numbers, the de- pixels, i.e., upper-left, upper-right, lower-left, and lower-right
signed spatial prediction scheme is processed by the following neighboring pixels, as (3)
steps:
Initial:
Let represent the value of the pixel ( , ) in an image.
(3)
Set variable as 1.
Step 1: Then, the in-place replacement is also performed in the pre-
As indicated in Fig. 4(a), the light-gray grids represent the dicted pixels by the value of ( ).
pixels marked as numbers , and the pixels marked as the num- Step 3:
bers bigger than are represented as white grids. The light-gray If the prediction process is complete, i.e., every pixel marked
grids denote the predicted pixels. Each arrow direction shows as nonzero is predicted, end the prediction process. Otherwise,
the corresponding pixel predicted by their cross neighboring let become and go back to Step 1. Note that the pixels
pixels, i.e., upper, left, right, and lower neighboring pixels, as that are treated in next iterate process are the pixels that have
(2) not been predicted yet, namely, the pixels that are represented
by white grids in Fig. 4(b). As the previous algorithm that marks
(2) the numbers, one fourth of all pixels in the present round are
treated as the new small image in the next iterate process.
Then, the value of is substituted by ( Note that this procedure is an in-place process, i.e., the orig-
). inal pixel values are replaced by the prediction errors. No extra
Step 2: memory space is required for storing the prediction error. Also
As depicted in Fig. 4(b), the dark-gray grids represent the notice that the prediction order is arbitrary in both Step 1 and
pixels marked as number and denote the predicted pixels Step 2 of the prediction algorithm because the predicted pixels
in this step. They are predicted by their diagonal neighboring marked as the same number do not refer to each other. Thus, a
KUO et al.: AN ESPIC 853

building the hierarchical structure, the ability of recovery in the

decoding phase and computational simplicity at the same time.
It is clear that there is migration behavior of errors. When
we use four rounds, the error of the prediction error for each
pixel in level 1 only affects itself. Nevertheless, the error of the
prediction error for each pixel in level 2 affects itself and eight
pixels in level 1 around the sample pixel. Similarly, the error of
the prediction error for each pixel in level 3 affects 49 pixels
(1 for self, 8 for pixels in level 2, and 40 for pixels in level 1)
which are all around the sample pixel. Furthermore, the error of
the prediction error for each pixel in level 4 affects 225 pixels (1
8 40 176) and the pixels marked as 0 affect 961 pixels. To
Fig. 5. Hierarchical structure tree.
sum up, the error of the prediction error for each pixel in level
affects ( ) pixels. As a result, the higher level pixels
sample of the relationship map among pixels is shown in Fig. 3. affect more pixels so that the higher level pixels have higher
The numbers in the map also express the hierarchical structure significance. This feature is valuable to the quantization phase
of an image. Note that all pixels marked as 0 may be marked as that is discussed in the next section.
9, 10, 11, 12, and so on if further prediction is carried out. Also
notice that the relationship map must be expanded to the whole
IV. SPIHT ALGORITHM AND ITS EXTENSION WITH
image. In other words, it is not restricted in a 16 16 or 32
MULTILEVEL THRESHOLDS
32 block.
For instance, if we use four rounds to predict, we can discover To further enhance compression performance, we also
that there are four pixels marked as 7 and four pixels marked as 8 modify the SPIHT algorithm [2] as our quantization scheme.
around every pixel marked as 0; moreover, every pixel marked The inputs of the original SPIHT algorithm are DWT coeffi-
as 8 has four surrounding pixels marked as 6, and every pixel cients. In our scheme, the hierarchical structure among pixels
marked as 7 has four surrounding pixels marked as 5. The rest in the image is established as we mentioned. This hierarchy
of the pixels are deduced by analogy. As a result, we can treat we built is very similar to the hierarchy made by DWT. Thus,
the prediction error image as several trees. Every tree has one combining our prediction scheme and SPIHT algorithm is
root, which is the pixel marked as 0, and the pixels marked as 1 very natural and smooth. This feature puts us in a position to
and 2 will be the leaves of the trees. The hierarchical structure integrate this hierarchy with many quantization methods of
of the prediction error image is presented as a tree, illustrated in wavelet or subband image-compression technologies. Note
Fig. 5. that many of the other quantization methods of wavelet and
In the prediction phase, all pixels marked as odd numbers and subband image-compression techniques are also available since
even numbers are executed in Steps 1 and 2 respectively except they are based on the hierarchical structure of the transformed
the pixels marked as 0. In the first round, the pixels marked as coefficients.
1 and 2 are predicted. In the second round, the pixels marked as One of the advantages of using the DWT quantization
3 and 4 are predicted, and so on. The remaining pixels, marked scheme, such as SPIHT or others, is the ability of generating an
as 0, are posed as the roots of the trees. However, the root pixels embedded bit stream; i.e., it can produce a progressive codec.
are never predicted but the quantity of the roots is quite small. Depending on available storage space or network bandwidth,
Note that each pixel uses four neighboring pixels to predict, one can use progressive codec to gain any bits of coded data.
and the distance is from every neighboring pixel to the The key reason for the ability of the progressive coding is that
central pixel, while denotes the level number in prediction the significant information has high quantization priority. While
processes. For instance, the pixels in level 1, pixels of number more significant information from images is ordered as the
1 and number 2, use the four nearest surrounding neighboring preceding position, the decoder can provide better quality from
pixels for prediction and the distance from every neighboring less coded data and stop at any number of bits when the space
pixel to the central pixel is 1. The pixels in level 2, pixels of or bandwidth is exhausted or the image quality is adoptable.
number 3 and 4, use four surrounding neighboring pixels for The progressive encoding characteristic is also realized in our
prediction and the distance from every neighboring pixel to the proposed scheme.
central pixel is 2. Similarly, the pixels in level 3 and 4 use four We extend SPIHT with multilevel thresholds. The quantiza-
surrounding neighboring pixels for prediction and the distance tion with multilevel thresholds takes the significance of predica-
from every neighboring pixel to the central pixel is 4 and 8, re- tion level into account, while quantization only depends on the
spectively. Notice that the level number of every pixel marked as absolute values of the transformed coefficients with the iden-
0 is the round number plus 1. On the other hand, in the decoding tical thresholds for each round in the traditional quantization
phase, the above process can be performed inversely without any approach of wavelet compression scheme. The traditional quan-
loss if no information is lost in the quantization process. This tization approach tends to ignore the significance of hierarchical
is also the key we can use information from the pixels behind level mentioned in the last paragraph of Section III. However,
the sample pixels. However, the reason for using the prediction different levels represent another dimension of importance. As
methods mentioned here is that we consider the potentiality of we mentioned, the predictive errors in higher level have more
854 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 12, NO. 10, OCTOBER 2002

(a)

(b)

(c)
Fig. 6. Relationship between multilevel thresholds and prediction errors.

effects than in the lower level. In our approach, we take this

into account in the determination of significance. In our experi- (d)
ence, the multilevel thresholds can be applied not only to SPIHT
but also to other quantization schemes. However, we improve
the SPIHT algorithm by modifying the definition of the signifi-
cance function , which indicates the significance of a set
of pixels from (4) and (5) (e)

if
(4)
otherwise
(f)
if
(5)
otherwise

and by modifying the initial value by (6) within the algorithm

II of SPIHT denoted in the original paper [2]. ( represents the
partitioning subset and represents the transform coefficient Fig. 7. Images and their energy diagrams before and after prediction
at coordinate ( , ). To save space in publication, please refer to = =
processes. (E entropy, 3D standard deviation.)
[2] to see the detailed notation.) We have
Fig. 7(a), (c), and (e). After applying four iterations of predic-
maxlevel of prediction (6) tion processes of ESPIC algorithm, the prediction-error images
and the corresponding energy diagrams are shown in Fig. 7(b),
(d), and (f). These figures clearly indicate that ESPIC algorithm
The multilevel thresholds used in the modified quantization are can converge the energy.
plotted in Fig. 6. The latitude axis denotes the prediction level,
Comparison using peak signal-to-noise ratio (PSNR) be-
and the longitude axis represents prediction errors. In the first-
tween ESPIC with our modified SPIHT and JPEG at different
order quantization in Fig. 6, we quantize only the level 4 pixels
compression ratios is shown in Fig. 8. The statistical charts in
whose absolute values of predicted errors are larger than 256.
Fig. 7 show that the performance of the proposed scheme is
Then, in the second-order quantization, both the level 4 pixels
very close to the baseline JPEG which uses DCT. Note that the
whose absolute values of predicted errors are larger than 128
compression ratios on the latitude axis of Fig. 8 are not regular
and the level 3 pixels whose absolute values of predicted errors
because we could not easily get just any compression ratio to
are larger than 256 are quantized. A pixel in level 4 with an
encode an image with JPEG, because the JPEG compression
error value above 128 has the same significance as a pixel in
scheme is not progressive.
level 3 with an error value above 256. The rest can be deduced
by analogy. In addition, it is well known that the PSNR does not perfectly
represent the compression quality; that is to say, it is not ac-
tually fair to use PSNR to compare two compression schemes.
V. EXPERIMENTAL RESULTS AND COMPARISONS
Thereby, we depict the subjective visual quality of ESPIC and
We use some examples to illustrate the performance of the JPEG in Fig. 9. For the Lena image, Fig. 10(a)–(f) represents
proposed ESPIC in this section. The original Lena, Barbara, and the compression ratio of 1 : 30, 1 : 61 and 1 : 86 by ESPIC and
Baboon images (512 512, 256 gray level) and their corre- JPEG respectively. Similarly, Fig. 9(g)–(l) and (m)–(r) present
sponding energy diagrams (gray-level histogram) are shown in the comparison of the Barbara and Baboon images respectively.
KUO et al.: AN ESPIC 855

Fig. 8. Comparisons of compression ratio.

It is obvious that JPEG brings the blocky effect at that high

compression ratio and ESPIC brings the smooth effect at the
same high compression ratio. In order to show these effects in
depth, the images of Barbara that are encoded and decoded via
ESPIC and JPEG at the 1 : 62 ratio are enlarged and shown in
Fig. 10. Although ESPIC is beaten by JPEG at some specific
compression ratio with PSNR measure, ESPIC still has better
visual quality for some special applications in subjective judg-
ments.
In the computational complexity issue, the ESPIC has the
clear advantage. No matter how many rounds we use in the
prediction phase, every pixel only needs four integer additions
and one bit-shift for all predicted pixels except the pixels
marked as “0” that are never predicted. We made a comparison
between ESPIC and Lifting Wavelet 5/3 [7], which needs much
less computational power than that of the other wavelet com-
puting approaches. The Lifting Wavelet Transform separates
all serial inputs into odd and even parts and these two parts play Fig. 9. Visual comparisons of decoding results between ESPIC and JPEG.
different roles to complete the transforming. For 5/3 Lifting
Wavelet Transform, the pixels of the odd part need two integer
VI. CONCLUSIONS
additions and two bit-shifts, and the pixels of even part need
two integer additions and only one bit-shift. Moreover, 2-D An efficient and progressive spatial prediction-based
wavelet transforms need to convolute the image vertically and image-compression scheme is developed in this paper. Al-
horizontally. As a result, the 5/3 Lifting Wavelet Transform though JPEG2000, which makes use of DWT as the core
needs four integer additions and three bit-shifts per pixel for technology, does achieve significant compression efficiency,
one scale transform. The result is demonstrated in Table I. and though a spatial prediction-based scheme may not be
However, the wavelet transform for image compression needs so attractive in image compression at the present time, we
more than one scale transform. successfully extend the 2-D spatial prediction-based scheme
856 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 12, NO. 10, OCTOBER 2002

improvement. However, this remains future work, to include all

of the above features in the spatial predication-based approach.

REFERENCES
[1] J. M. Shapiro, “Embedded image coding using zerotrees of wavelet co-
efficients,” IEEE Trans. Signal Processing, vol. 41, pp. 3445–3462, Dec.
1993.
[2] A. Said and W. A. Pearlman, “A new fast and efficient image codec
based on set partitioning in hierarchical trees,” IEEE Trans. Circuits
Syst. Video Technol., vol. 6, pp. 243–250, Jun. 1996.
[3] H. J. Wang, Y. L. Bao, C.-C. Jay Kuo, and H. Chen, “Multi-Threshold
Wavelet Codec (MTWC),” in Proc. JPEG 2000 Convergence Phase
Meeting, Geneva, Switzerland, Mar. 1998, ISO/IEC JTC1/SC29/WG1
N819.
[4] W. C. Huang and L. W. Chang, “Predictive subband image coding with
wavelet transform,” Signal Processing: Image Commun., vol. 13, pp.
171–181, 1998.
[5] R. C. Gonzalez and R. E. Woods, Digital Image Processing. Reading,
MA: Addison-Wesley, Sept. 1993, pp. 358–361.
[6] W. B. Pennebaker and J. L. Mitchell, JPEG Still Image Data Compres-
sion Standard. New York: Van Nostrand, 1993.
[7] I. Daubechies and W. Sweldens, “Factoring wavelet transforms into
lifting steps,” J. Fourier Anal. Applic., vol. 4, pp. 247–269361, 1998.

Chin-Hwa Kuo (S’93–M’00) received the B.S.

Fig. 10. Enlarged images of the visual comparison between ESPIC and JPEG degree in mechanical engineering from Chung-Yuan
(ratio= 1 : 62). University in 1980, the M.S. degree in electrical
engineering from Marquette University, Milwaukee,
WI, in 1989, and the Ph.D. degree in electrical
TABLE I engineering from the University of Notre Dame,
COMPARISON OF COMPUTATION COST BETWEEN ESPIC AND South Bend, IN. in 1994.
LIFTING WAVELET 5/3 Currently, he is an Associate Professor in the De-
partment of Computer Science and Information Engi-
neering, and Director of the Division of Distance Ed-
ucation, both at Tamkang University, Taipei, Taiwan,
R.O.C. His research interests include multimedia processing and applications,
computer-assisted learning, and adaptive systems.
Dr. Kuo is a member of the IEEE Circuits and Systems Society’s Technical
Committees on multimedia systems and applications.

into the lossy image-compression domain and thus retain its

Tzu-Chuan Chou received the B.S. degree in com-
simplicity. Furthermore, the developed multilevel threshold puter science and information engineering in 1996
scheme takes into account the significance of the pixel values and the M.S. degree in computer science and infor-
and the hierarchical levels to further enhance the compression mation engineering in 1998, both from Tamkang Uni-
versity, Taipei, Taiwan, R.O.C., where he is currently
performance. Our design and simulation results indicate that the working toward the Ph.D. degree in the Department
developed ESPIC scheme consists of the following significant of Computer Science and Information Engineering at
features: progressive coding, high computation efficiency, less Tamkang University.
His research interests include multimedia informa-
memory requirement, and good visual quality and compression tion retrieval, image compression, image processing,
efficiency. As a result, the developed scheme is especially and web-based integrated applications.
beneficial to the applications of real-time and wireless image
transmission under limited bandwidth and computation power
environments. Tay-Shen Wang received the B.S. degree in com-
The methodology we chose in the predication phase estab- puter science and information engineering in 1998
and the M.S. degree in computer science and infor-
lishes a hierarchical structure among pixels in the encoding mation engineering in 2000, both from Tamkang Uni-
stage. The approach taken here also gives consideration to the versity, Taipei, Taiwan, R.O.C.
potentiality of building the hierarchical structure, the ability of He is currently with Cyberlink Corporation,
Taiwan, R.O.C. His research interests include
recovery in the decoding phase and computational simplicity. surveillance system, image processes, and video
Therefore, the filters selected in this paper do not guarantee to coding technologies.
be optimal. Better selection of filters may result in performance