A N EFFICIENT VISUAL PATTERN BLOCK TRUNCATION CODING
Liang-Gee Chen and Yuan-Chen Liu
D S P / I C Design Lab., Department of Electrical Engineering National Taiwan University, Taipei, Taiwan, R.O.C.
Email: [email protected]
ABSTRACT
This paper presents a new practical image coding sys- tem, visual pattern block truncation coding ( V P B T C ) . T h e first idea of this new algorithm is the application of block truncation coding ( B T C ) , which encodes a n original image and obtains a bitmap. This bitmap is applied to computing t h e block gradient orientation and matching the block pat- tern. Another refinement is t h a t the classification of blocks are based on properties of human visual perception. Blocks are classified into two types: uniform blocks and edge block- s. Not only can V P B T C remove the drawbacks of B T C (e.g. higher bit rate) but also can i t repair the defects of VPIC (e.g. mismatching p a t t e r n ) . Four benchmarks prove t h a t the MSE of this algorithmis less t h a n t h a t of V P I C by 37%. Its bitrate, furthermore, is about 60% of BTC’s. T h e ad- vantages of this coding method are fine edge preservation, low computational complexity, and easy implementation.
1. INTRODUCTION
Image coding is getting significant for current applications, for instance, H D T V , multimedia, videophone, video con- ference and video storage. To remove possible redundan- cy i n t h e related d a t a is indispensable since there are nu- merous d a t a t o be processed. T h e most efficient method is d a t a compression. Spatial and/or temporal redundancy are eliminated t o accommodate communication or storage. T h e common techniques for d a t a compression are predictive coding, transform coding, vector quantization ( V Q ) , visual pattern image coding ( V P I C ) , and block truncation coding ( B T C ) , etc.
B T C is a n efficient method for block correlated signal- s. I t is a moment preserving quantizer [1]-[6], t h a t is, it quantizes block d a t a into two values by preserving the first moment and t h e second moment. T h e advantages of B T C are simple computation, easy implementation and fine edge preservation, etc. T h e main problem of B T C is i t s high bitrate. I n a fixed block size B T C , the bitrate of B T C is about 1.625 bits per pixel. There are some algorithms [7, 81
which can obtain a lower bitrate and fewer error. However, t h e computation is complex so t h a t it is hard t o implement i n VLSI design and in real time processing.
V P I C is a high quality algorithm introduced by P . Chen and A. Bovik [9]. Unlike other algorithms, VPIC uses pix- el values i n a block t o compute block gradient orientation and t o select the bilevel block pattern with simple viewing geometry model. Pixels are quantized into two levels by the block mean, gradient magnitude, gradient orientation, and predefined block pattern. High compression ratio is a
considerable merit of VPIC. Nevertheless, if a block does
0-7803-3073-0/96/$5 .OO O1996 IEEE
770not contain an obvious edge, V P I C will mismatch the block pattern. Since V P I C uses the gradient magnitude t o clas- sify uniform blocks, it will make serious error when a block contains an edge in a low gray level background.
A new algorithm, visual pattern block truncation cod- ing ( V P B T C ) , is proposed in this paper. B T C is used t o encode an original image. This algorithm defines the edge block according to human visual perception. If t h e differ- ence between t h e two quantized values of B T C is larger than a threshold which is defined by visual characteristics, it will be defined as an edge block. In an edge block, the bitmap is adapted to compute block gradient orientation and to match the block pattern. Experimental results show t h a t t h e MSE of t h e algorithm is less t h a n V P I C by half in the same bitrate. This coding system has many merits such as fair edge preservation, low computation complexity, high compression ratio, easy implementation, and suitable reconstructed image.
In this paper, section 2 briefly reviews B T C and VPIC algorithm. Section 3 discusses how to define t h e pattern of
a block. V P B T C algorithm is presented in section 4. Some simulation results are shown in section 5. Conclusions are given in section 6 .
2. P R E V I O U S A L G O R I T H M REVIEW 2.1. Block Truncation Coding
Basic B T C based on moment preserving is a bilevel quan- tizer. In the original coding scheme, an image is divided into n pel by n line blocks (typically n = 4). Basic B T C uses the mean of the pixel values in a block as the threshold
,
X t h . T h e basic algorithm computes two quantized values, andYI,
by preserving t h e first moment and the second moment in each block. T h e quantized values ofYn
andYI
can be defined asr--
where
5?
is the mean of pixel values in a block; U is t h estandard deviation of the pixel values in a block; a is the number of
X,
which is greater t h a n X t h ; ,B is the numberof X ; which is less t h a n or equal to Xth.
A bitmap records each pixel which belongs to an alter- native quantized values. T h e bitmap, t h e mean, and the standard deviation need to be transmitted. T h e bitrate of
basic B T C is 2 bits/pixel. When a two-dimension coding scheme is used [l], the bitrate can be reduced to 1.625 bit- s/pixel. T h e coding process of basic B T C takes only a few computation steps. Since basic B T C algorithm bases on p- reserving statistical moments, the quality of reconstructed image is commendable. Although holding manifold advan- tages, the main problem of B T C is its low compression ratio. As a result, there is interest in finding a fast algorithm of high compression ratio.
Lena 4 x 4 baboon 4 x 4 pepper 4 x 4 j e t 4 x 4 2.2. Visual Pattern Image Coding
V P I C is a high uality algorithm which adapts the pattern of subimage w i t 1 simple viewing geometr model. T h e o- riginal image bein divided into 4 x 4 bloczs, the two direc- tional variations, i x b i , j and Aybi,j, can be computed from
A,bi>, = A V E ( I , , , : 4i
+
2 5 n 5 4i+
3 , 4 j 5 m 5 4 j+
3 ) - A V E ( I , , , : 4i5
n5
4i+
1 , 4 j 5 m5
4 j+
3) A y b i , , = A V E ( I , , , : 4 i 5 n 5 4i+
3 , 4 j+
2 5 m 5 4j+
3) - A V E ( I , , , : 42 5 5 42+
3 , 4 j 5 m 5 4 j t 1) ( 3 ) ( 4 )T h e block gradient magnitude (lAbli,j) and block orienta- tion (LAbi,j) can be found by t h e following equations
IAbli,j = J(Axbi,j)’
+
(Aybi,j)’ (5)V P I C B V P I C - ~ 1 0 0 %
127.58 83.77 65.66%
634.62 388.71 61.25%
185.19 107.11 57.84%
279.73 148.04 52.92%
If a block gradient magnitude is larger t h a n a certain thresh- old, i t will be defined as a n edge block. VPIC then uses the block gradient orientation t o map a block pattern. If it is less t h a n the threshold, the block will be identified as an uniform block. Whenever a block is an uniform block, the encoder will transport the block mean and a header. By contrast, if a block is a n edge block, the encoder will trans- mit t h e mean, the gradient magnitude, and the index of the block pattern.
T h e decoder is able to reconstruct images easily by trans- mitting block pattern, gradient magnitude, header, and block mean. T h e essential benefit of VPIC is t h a t it cata- logues blocks according t o human optical perception. V P I C also has the advantages of high compression ratio, fine edge preservation and low computation complexity. Pos- sessing such merits, V P I C still requires efforts to guaran- tee a more practical solution under several subjective cri- teria. A chief defect of V P I C is t h a t whenever the edge of a block is not distinct, the algorithm would frequently mismatch t h e block pattern. Taking figure 1 as example, Axb,,,=3/8, Ayb,,,=9/8. T h e result of t h e block gradien- t is computed as 90’. From the bitmap, we can see the correct block orientation is -45’. I t obviously mismatches the block pattern. T h e example manifests t h a t it is sig- nificant t o find a better VPIC algorithm under subjective criteria.
3. BLOCK PATTERN DECISION
B T C is t h e most approvable algorithm if t h e drawback of low compression ratio is remedied. This new B T C based algorithm applies VPIC t o reduce the bitrate of BTC. To improve the defects of mismatching patterns by VPIC we uses t h e bitmap of B T C t o select a block pattern. ?he
rocess is in the following. First, an image is s lit into 4 x 4 glocks.
& ,
and the bitmap are obtained a i e r B T C has encoded a 4 x 4 block. Utilizing the coded bitmap, we find97 97
98
101
BTC
0 0 0
I
95
95 97 98
[--2
0
0 0
0
105 105 95 97
-/
I
1
0 0
99 99 103 97
1
1
1
0
Figure 1. An Example of the mismatching pattern of VPIC.
A x and A y which represent the gradient of x direction and
y direction.
A, = S U M ( B , , , : 42
+
2 5 n 5 4i+
3 , 4 j 5 m 5 4 j+
3) - S U M ( B , , , : 42 5 n 5 42+
1 , 4 j 5 m 5 4j+
3 )A,, = S U M ( B , , , : 4i 5 n 5 4 i + 3 , 4 j + 2 5 m 5 4 j + 3 ) - S U M ( B , , , : 4i 5 n 5 42
+
3 , 4 j 5 m 5 4 j+
I) ( 8 )Where Bn,,,, is the bitmap of the block. T h e block gradient orientation can be found out by the following equation
(7)
(9) After computing the block gradient orientation, we can compare the bitmap with t h e previously defined block pat- tern. An index of the most similar block p a t t e r n is taken as the index of bitmap.
F o u r benchmark images are taken as simulation. T h e d- T A B L E I
THE COMPARISONS O F V P I C AND B V P I C
imaqe Block size
I
M S E RatioFigure 2. Contrast sensitivity measurement.
4. VISUAL PATTERN BLOCK TRUNCATION
Two types of blocks, uniform blocks and edge blocks, are usually classified by using block variance. There are several methods of defining blocks. We propose t h a t the identifica- tion of a block should base on t h e characteristics of human vision. After t h e encoding process of B T C , there are two quantized values whose difference represents the changes in t h e intensity of illumination of a block. According to t h e Weber-Fechner Law [ l o ] , t h e quotient of the fraction of il- luminative intensity i n a region is an essential reference for h u m a n visual perception. T h e sensible stimulus for human eyes t o the changes of illumination, the contrast sensitivity, is dependent on t h e intensity of the surrounding. As figure 2 shows, a given patch of light of intensity I + A I surrounded
by a background of intensity I .
T h e sensible difference for normal human eyes A I is to
be determined as a function of I. T h e ratio A I / I , known as
Weber fraction, is almost constant a t the value about 0.02 [lo]. Recognizing t h e value of minimum contrast sensitivity of h u m a n eyes, we define t h e block whose difference of illu- minative intensity is perceivable as a n edge block and the imperceptible one as an uniform block. Let the difference between
l’i
a n d&
be A I , andYO
be t h e background I. If the quotient ofl’i
-Y
o
and y0 is less t h a n 0.02, the block is an uniform block. Otherwise, i t is an edge block. T h e flowchart of V P B T C is shown in figure 3 and its procedures are described as follow:1. An image is divided into n pel by n line blocks (typi- 2. T h e block is coded by B T C .
3 . If t h e ratio of
YI
-YO
andYO
is less t h a n 0.02, it is an 4. If i t is edge block, the bitmap is used t o decide the 5. T h e quantized values or mean are transmitted and soCODING
cally n = 4).
uniform block. Otherwise, it is an e d g e block. block pattern by bitmap based VPIC. is the index of the block pattern.
5 . SIMULATION RESULTS
Four benchmarks are used as test examples. Table I1 shows t h e results of various algorithms. T h e bitrate is fixed in about 0.9 bit per pixel. V P B T C ( 1 ) takes t h e block mean as
the threshold and
Yo, YI
as t h e quantized values of momentpreserving. V P B T C ( 2 ) uses mean as the threshold of B T C .
E;)
andYI
are set as the lower mean and t h e upper mean re- spectively. V P B T C ( 3 ) is O B T C presented by Chen and Liu[E]. T h e reconstructed image of new algorithm 2 is shown in figure 4.
6 . CONCLUSION
This paper proposes a new developed image coding system, V P B T C , which not only removes the drawbacks of B T C and V P I C b u t also maintains their advantages. T h e sys- tem surpasses the V P I C in accuracy, arithmetic simplicity and hardware implementation. Furthermore, we also offer an visual based B T C algorithm which outdoes B T C in bi- t r a t e and image quality. Four benchmark images are tested to verify the performance. T h e experimental results show t h a t t h e proposed system are superior t o other algorithms. In addition, the regularity and simplicity of t h e V P B T C algorithm manifest t h a t it is quite suitable for an efficien- t VLSI implementation when used in a low bit r a t e video coder. A systolic architecture on real video codec system is currently under development for real-time application.
REFERENCES
[I] E. J. Delp and 0. A. Mitchell, ”Image compression us- ing block truncation coding,” I E E E Trans. Commun.,
vol. COM-25, no. 9 , Sept. 1979, pp. 1335-1342. [ 2 ] 0. R . Mitchell and E. J . Delp, ”Multilevel graphics
representation using block truncation coding,” Proc. I E E E , vol. 6 8 , no. 7, July 1980, pp. 868-873.
[3]
D.
R. Halverson, N . C. Griswold, and G. L. Wise, ” A generalized block truncation coding algorithm for im- age compression,” I E E E Trans. Acoust., Speech, Sig- nal Processing, vol. ASSP-32, no. 3, J u n e 1984, pp.6 64-6 68.
[4] L. G. Chen, Y. C. Liu, T .
D.
Chiueh, a n d Y. P . Lee, ” A real-time video signal processing chip,” I E E E Trans. Consumer Electronics) vol. 39, no. 2 , may 1993, pp. 82-92.[5] T . P . Chan, B. Zeng, and M. L. Liou, ” B T C image cod- ing with visual patterns,” in Proc. APCCAS’94 Conf.,
Grant Hotel, Taipei, Dec. 1994, pp. 12C.8.1-12c.8.6. [ 6 ] V. R . Udpikar and J . P. Raina, ”Modified algorithm
for block truncation coding of monochrome images,”
Electron. Lett., vol. 2 1 , no. 20, Sept. 1985, p p . 900-902. [7] M. Kamel, C. T. Sun, and Lian G u a n , ”Image compres- sion by variable block truncation coding with optimal threshold,” I E E E Trans. Signal Processing, vol. 39, no.
1, January 1 9 9 1 , pp. 208-212.
[E] L. G. Chen and Y. C. Liu, ” A high quality MC-BTC codec for video signal processing,” I E E E Trans. CZT- c u d s a n d S y s t e m s for V z d e o T e c h n o l o g y , vol. CSVT-4, no. 1, Feb. 1994, pp. 92-98.
[9] D. Chen and A. C. Bovik, ”Visual P a t t e r n Image Cod- ing,” I E E E Trans. Commun., vol. COM-38, no. 1 2 , Dec. 1 9 9 0 , pp. 2137-2146.
[ l o ] W . K . P r a t t , Dzgital Image Processzng. New York:
John Wiley & Sons, Inc, 1 9 9 1 .
Original Image
baboon pepper j e tI
No
20.11 21.94 22.50 22.52 25.45 30.72 31.24 31.33 23.66 29.44 29.86 30.22Bitmap Based
VPIC
1
Mean
t
Index of pattern,
Yl , Yo
Figure 3. Flowchart of VPBTC. T A B L E I1 T H E P E R F O R M A N C E O F V P B T C 58.31 52.29 51.02 365.88 364.29 55.13 48.93 47.84 279.81 74.00 67.14 61.84 ( a ) M S E ImageI
V P I C V P B T C ( 1 ) V P B T C ( 2 ) V P B T C ( 3 ) Lena I 5.96 4.21 4.05 3.93 baboon 16.29 12.95 12.14 11.84 pepper 6.61 4.34 4.11 3.99 iet1
7.50 3.86 4.04 3.86 ( b ) M A E ImageI
V P I C V P B T C ( 1 ) V P B T C ( 2 ) V P B T C ( 3 ) LenaI
27.07 30.41 30.95 31.05Figure 4. Result of VPBTC algorithm. (a) Origi- nal image is 5 1 2 x 5 1 2 pixels with 8 bits gray level resolution; (b) Coded image using VPBTC(2).
