Design of Odd Number Rational Coefficients Symmetric Compactly-supported Biorthogonal Wavelet Filters Han Fang-fang Duan Fa-jieˈDuan Xiao-jieˈZhang Chao State Key Lab of Precision Measuring Technology & Instruments, Tianjin University Tianjin, 300072, China School of Electrical Engineering, Tianjin University of Technology Tianjin, 300382, China fangfanghan2004@163.com State Key Lab of Precision Measuring Technology & Instruments, Tianjin University Tianjin, 300072, China fajieduan@126.com Abstract—Most wavelet filters that constructed based on Daubechies compactly-supported wavelet construction theory have nonlinear phase and irrational number coefficients. Nonlinear phase may cause distortion in image processing, and irrational filter coefficients will cause much inconvenience for wavelet applications on computer, especially for embedded processor applications. With thought of the theory of perfect-reconstruction filter banks, this paper does a study for linear phase biorthogonal perfect-reconstruction wavelet filter banks, and through of detail derivation, makes a conclusion for the procedure of odd number rational coefficients symmetric compactly-supported biorthogonal wavelet filter construction, which can be achieved by adding some vanishing moments conditions. With the example of length 7-5 biorthogonal wavelet filters design, this paper makes a detail discussion on the value selection for free variable appears in design equations, and then deduces the length 7-5 rational biorthogonal wavelet filters models. On the other hand, most biorthogonal wavelet filters’ coefficients obtained based on Daubechies methods are irrational numbers [1, 2], which cause much inconvenience in wavelet applications especially on computer hardware operation. Therefore, since wavelet transform can be seen as a special structure of perfect-reconstruction filter bank operation, we can add some proper conditions, such as vanishing moment constraints, to construct a rational coefficients biorthogonal wavelet filter bank. II. CONSTRUCTION OF BIORTHOGONAL PERFECT-RECONSTRUCTION WAVELET FILTER BANK Figure 1 shows a group of biorthogonal perfectreconstruction wavelet filters [3, 4]. The upper branch is constructed by two low-pass filters, where H (z ) is ~ decomposition low-pass filter, and H ( z ) is reconstruction low-pass filter. The lower branch is constructed by two high-pass filters, where G (z ) is decomposition high-pass ~ filter, and G ( z ) is reconstruction high-pass filter. p 2 Keywords-wavelet filter construction; rational coefficients wavelet filter; odd number symmetric filter; biorthogonal wavelet; wavelet transform I. represents down-sampling by 2, and up-sampling by 2. INTRODUCTION Wavelet multi-resolution analysis is performed under the orthogonal conditions. But when to solve the practical engineering problems, only orthogonal condition is inadequate. For example, in order to avoid distortion in image transformation processing, the response functions of filters came from wavelet two-scale functions preferably have linear phases. The necessary and sufficient condition for filter’s linear phase is that the impulse response of filter must be symmetric about its center point. But according to the orthogonal wavelet transform theory and its construction methods under orthogonal multi-resolution thoughts, except Haar wavelet, the other wavelets’ low-pass filter coefficients are not symmetric, i.e., have no linear phase [1, 2]. In order to overcome this limitation, in virtue of the perfect-reconstruction filter banks thoughts, the filters for signal decomposition and reconstruction can be not the same one, and then a biorthogonal wavelet transformation system based on biorthogonal structure be constructed. H (z ) x(z ) p 2 n 2 n 2 represents ~ H ( z) y (z ) G (z ) p 2 n 2 ~ G( z) Figure 1. Structure of biorthogonal perfect-reconstruction wavelet filter bank If there is no the middle procedure of signal processing, the output of perfect-reconstruction filter bank should be a perfect reconstruction of the input, or a perfect reconstruction with some time delay. Then the four filters ~ ~ of H (z) , H ( z ) , G (z ) and G ( z ) should meet the following relationship, where F0 ( z ) is called distortion term, and F1 ( z ) is called alias term[3, 4]. 76 F0 ( z ) F1 ( z ) ~ ~ H ( z ) H ( z ) G ( z )G ( z ) 2 z l ~ ~ H ( z ) H ( z ) G ( z )G ( z ) 0 Assign < as follow equation: (1) 2 (2) < As long as the four filters meet the requirements represented by equation (1) and equation (2), they can construct a perfect-reconstruction filter bank. Therefore, filters can be designed freely according to those relationships. Here, this paper proposes the following filter design way: ~ ~ ­°G ( z ) zH ( z )ˈviz. G ( z ) zH ( z ) (3) ®~ °̄G ( z ) z 1 H ( z ) N 1· 2 § K 1 ¨ < ¸ Q(< ) 2¹ © H (< ) (7) ~ N 1· 2 ~ § K 2 ¨ < ¸ Q (< ) 2¹ © ~ H (< ) (8) Therefore: ~ H (< ) H (< ) Y (< ) ~ NN 1· 2 § ¨< ¸ 2¹ © ~ Q(< )Q (< ) (9) Y (< ) should meets two requirements: ķ the perfect-reconstruction condition represented by equation (5); ĸDaubechies compactly-supported condition. First, for the perfect-reconstruction condition, thinking combination with equation (5) and equation (6), and considering with the filter symmetric feature, Y (< ) should meets the follow equation: Then: 2 (6) Then: Take equation (3) into equation (1) and equation (2), and order the time delay l 0 , then F1 ( z ) 0 , F0 ( z ) 2 . Assign: ~ Y ( z) H ( z) H ( z) (4) Y ( z ) Y ( z ) § 1 z 1 · ¸ 1 z¨ ¨ 2 ¸ 2 © ¹ (5) Therefore, for the four perfect-reconstruction filters, we can firstly design the two low-pass filters H (z ) and ~ H ( z ) , and consequently deduce the two high-pass filters ~ G (z ) and G ( z ) according to equation (3). Hence, in the following, this paper will concentrate on the symmetric ~ compactly-supported filters design for H (z) and H ( z ) . Y (< ) Y (< ) 1 (10) Second, for the compactly-supported condition, suppose: ~ P(< ) Q(< )Q (< ) iZ Since z e , and combination with Euler's formula and Trigonometric Formulas, P (< ) can be represented by III. CONSTRUCTION OF ODD NUMBER SYMMETRIC COMPACTLY-SUPPORTED BIORTHOGONAL WAVELET FILTERS symbol x , where x sin 2 (Z / 2) : The original intention to construct biothogonal wavelet structure is to make the wavelet filter not only form a perfect-reconstruction system but also have linear phase. The necessary and sufficient condition for filter’s linear phase is that the impulse response of the filter must be symmetric about its center point, that is i 2OZ H (Z ) e H (Z ) , or represents by coefficients ~ P( x) Q( x)Q ( x), x sin 2 Z 2 With Taylor methods to get a group of special solution: P( x) relationship hk h2O k . When hk R , coefficients can be regarded as symmetric about the point of O . This paper’s purpose is to design odd number filter, so this paper proposes the rule that the designed filter is symmetric about its axis 0, that is, the filter’s coefficients meet equations of hk hk . Therefore, according to the filter design methods proposed from references [2]~[10], the two low-pass filters for odd number symmetric biorthogonal wavelet can be designed as follow procedure [2, 5-10]. Suppose the decomposition low-pass filteris represented by H (z) , and the reconstruction low-pass filter is ~ represented by H ( z ) . The vanishing moments of the two ~ filters are N and N respectively. ~ ( N N ) / 21 ¦ n 0 C (nN N~ ) / 21 n x n ~ N N x 2 r ( x) (11) r (x) in equation (11) should meet the equation: r ( x) r (1 x) 0 (12) What’s more, P(x) should meet the Daubechies compactly-supported condition: max x[ 0,1], r ( x ) r (1 x ) 0 77 P( x) ~· § N N ¸ 1 2¨¨ ¸ 2 ¹ 2 © (13) IV. DESIGN FOR LENGTH 7-5 RATIONAL COEFFICIENTS SYMMETRIC COMPACTLY-SUPPORTED BIORTHOGONAL WAVELET FILTER BANK To construct orthogonal compactly-supported wavelets based on Daubechies methods, if the vanishing moments takes the highest order value, that is, Q(1) z 0 and as ~ same as Q(1) z 0 , the constructed wavelet function and scale function will have a rapid attenuation speed, but the coefficients of these filters can not be rational. To get the rational filter coefficients, a method is to reduce the order of vanishing moments, and then make the designed equation appear one or more free variables. Select suitable value for free variables, thus obtain the rational coefficients wavelet filter [2, 5-10]. So the key problem for rational coefficients wavelet filter design is how to select free variables’ value in Daubechies compactly-supported restriction range. A 4 B 4 At 4 Bt t < B At 4 Bt < Bt ] To meet the equation (10), coefficients of < 4 and < in equation (16) should be zero, and constant term should be 1/2. Therefore, there is the equation group: ­ °4 A 4t 4 0 ° (17) ® A 4 B 4 At 4 Bt t 0 °K K 1 ° 1 2 Bt 2 ¯ 4 The solution for the equation group (17) with free variable t is: ­ ° A t 1 ° 4t 2 4t 1 ° (18) ®B 4t 4 ° 8t 8 ° ° K1 K 2 3 4t 4t 2 t ¯ Now discusses the compactly-supported condition. For P (< ) : ~ P(< ) Q(< )Q (< ) ~ H (< ) 1· § K 2 ¨ < ¸< t 2¹ © 2 P(< ) 4t 4t t 3 2 [4(t 1)< 3 4(t 1)< 2 (19) (4t 3 4t 2 1)< t (2t 1) 2 ] Take then: According to equation (7) and equation (8), and ~ N 2, N 2 , there is: K 1 K 2 (< 2 A< B)(< t ) Take solution equations (18) into P (< ) : B. Solving for P(x) H (< ) (16) 2 2 A. Selection for Wavelet Filter’s Vanishing Moment Suppose the low-pass filter of length-7 is H (z) , and its vanishing moment is N ; and the low-pass filter of length-5 ~ ~ is H ( z ) , and its vanishing moment is N . Because the vanishing moment for odd-length and symmetric biorthogonal filter must be even number, for the biorthogonal wavelet of length 7-5, the vanishing moments ~ ~ N and N are both even number, and N and N should be in such range: ~ 75 NN d 6 2 ~ Therefore, values for N and N have the following ~ ~ several probabilities: (1) N 2, N 4 ; (2) N 4, N 2 ; ~ (3) N 2, N 2 . For the first two combination probability values, there will be no free variables in wavelet construction equations. But for rational coefficients wavelet filter design, we need the free variables in design equations, so this paper prefers the third probability for vanishing ~ moment, that is N 2, N 2 . 1· § K 1 ¨ < ¸ < 2 A< B 2¹ © ~ H (< ) H (< ) K1 K 2 [4< 5 4 A 4t 4 < 4 4 4 A 4 B 4 At 4t 1< 3 Y (< ) z e iZ cos Z i sin Z 2 into equation (6), 1 § 1 z 1 · 1 ¸ 1 zz z ¨¨ cos Z ¸ 2 2 4 2 © ¹ Take upper < into equation (19), x sin 2 (Z / 2) , then achieve the equation: < (14) (15) P( x) 1 2 x Then: 4(t 1) t (2t 1) 2 x2 1§ 2 Z· ¨1 2 sin ¸ 2© 2¹ and assign 8(t 1) t (2t 1) 2 x3 (20) Because P(x) resolved by equation (20) must meet the Daubechies compactly-supported condition represented by equation (13), the value range for free variable t should be discussed. Try to select different value for t in its 78 D. Solving for Coefficients of Filters Combined equation (6) and equation (18), and rewritten equation (14) and equation (15), then: K1 3 8t 2 3t 1 H ( z) z [ z (4t 2) z 2 t 1 64 24t 2 20t 4 t 1 8t 2 3t 1 1 z (4t 2) z 2 z 3 ] t 1 K2 2 ~ H ( z) [ z (4t 2) z 16 (8t 2) (4t 2) z 1 z 2 ] permission range, we can do the rational coefficients wavelet filter design. C. Determination of Value Range for Free Variable t Rewritten equation (20) as the form as equation (11), then: 1 ¦ C1n n x n x 2 r ( x) P( x) n 0 ª 4(t 1) 8(t 1) 1 2x x 2 « 2 t (2t 1) 2 «¬ t (2t 1) º x» »¼ 4(t 1) 8(t 1) x , and it t (2t 1) 2 t (2t 1) 2 satisfies the condition of r ( x) r (1 x) 0 shown in equation (12). To meet Daubechies compactly-supported condition, P( x) should meet the requirement of equation (13), that is: r ( x) Therefore, max x[ 0,1], r ( x ) r (1 x ) ª 4(t 1) 8(t 1) 1 2x x 2 « 2 0 t (2t 1) 2 ¬« t (2t 1) In the value range of t , select different value for t , K 1 , K 2 , then we can construct different wavelet filters. When select a suitable value for t , it is possible to construct rational coefficients wavelet filters. Here gives an example. When t 3 / 2 , according to filter impulse response: º x» 23 ¼» H (Z ) (1 / 2) ¦ hk e ikZ , z e iZ kZ According to equation (20), plot the relationship curve figure about t l max P( x) . For every t , there is an equation And wavelet filter coefficients condition: ¦ hk about x l P(x) . What’s more, for x sin (Z / 2) , the value of x was limited in [0,1] . In the range of x [0,1] , calculate the max value of P(x) .Then for every t , exist 2 2 kZ It can be calculated that K 1 2 / 3 , K 2 1 . Then coefficients of the two low-pass filters are: an max P( x) , thus we can get the Figure 2. Noted that for equation (20), the denominator should not be zero, t (2t 1) z 0 . So the value of t is composed of three sects, that is (f,0.5) , (0.5,0) and (0,f) , shown as Figure 2. h k : h 3 h0 ~ ~ hk : h 2 1 1 25 , h 2 , h1 , 48 12 48 14 25 1 1 , h1 , h2 , h3 12 48 12 48 1 ~ 1 ~ 5 ~ 1 ~ 1 , h0 , h1 , h2 , h1 8 2 4 2 8 And according to equation (12), the high-pass filters are: g k : g 1 g~ k : g~ 4 Figure 2. Relationship of free variable t and g~ 1 max P( x) max The value range (f,0.755) (0.216,f) . P ( x) 8 for t 1 8 The scaling function and wavelet function constructed by this group of filters [4, 11] is shown in Figure 3. The amplitude-frequency characteristic and phasefrequency characteristic is shown in Figure 4. In figure 4, H1 represents the filter of H (z ) , H 2 represents the ~ filter of H ( z ) , G1 represents the filter of G (z ) , and According to Figure 2, when: x[0,1],r ( x ) r (1 x ) 0 1 1 5 1 , g 0 , g1 , g2 , g3 8 2 4 2 1 ~ 1 25 , g 3 , g~ 2 , 48 12 48 14 ~ 25 ~ 1 ~ 1 , g 0 , g1 , g 2 12 48 12 48 is 79 ~ G 2 represents the filter of G ( z ) . From Figure 4, we can see that the four filters all have linear phase. avoid image distortion in image transformation, so it is quite compatible for image processing. The difficulty for the rational wavelet filter design method proposed by this paper is solving for higher order and multivariable functions, and determination of the value range for the free variable under the effect of multiple parameters. This paper makes a discussion for length 7-5 biorthogonal wavelet filters, but when filter length is increased, the calculation difficulty will be increased more, which is need a much deeply study and discussion. REFERENCES [1] Daubechies I. Orthonormal basis of compactly supported wavelets. Variation on a theme. SIAM. Math. Anal, 1993, (24)2: 499-519. [2] Cheng Lizhi, Wang Hongxia, Luo Yong. Wavelet theory and applications. Science Press, 2006.10: 100-116 [3] Sweldens W. The lifting scheme: A custom-design of biorthogonal wavelets. Applied Computational and Harmonic Analysis, Volume 3, Issue 2, April 1996: 186-200 [4] Dwight F. Mix, Kraig J. Olejniczak, translated by Yang Zhihua, Yang Lihua. Tutorials for wavelet basis and application. Machinery Industry Press, 2006.4: 146-153 [5] Cohen A, Daubechies I, and Feauveau J C. Biorthogonal bases of compactly supported wavelets. Commun. Pure Appl. Math, 1992, 45(5): 485–560. [6] Kuang Zheng, Cui Minggen. Rational Filter Wavelets. Journal of Mathematical Analysis and Applications, Volume 239, Issue 2, 15 November 1999: 227-244 [7] Liu Zaide, Zheng Nanning, Song Yonghong, etc. Design of 9/7 biorthogonal wavelet filter with rational coefficients and high performance. Journal of Xi’an Jiaotong University, 2005.8: 847-851 [8] Liu Zaide, Zheng Nanning, Liu Yuehu, etc. Optimization design of 17/11 biorthogonal wavelet and its performance analysis for image compression. Journal of Electronics & Information Technology, 2007.6: 1403-1407 [9] Wang Hongxia, Cheng Lizhi, Wu Li. Construction for M-layer rational coefficients biorthogonal wavelet filter. Progress in Natural Science, Vol.13, No.2, 2003: 132-137 [10] Bilgin A, SementilliP J, Sheng F, et al. Scalable image coding using reversible integer wavelet transforms. IEEE Trans on Image Processing, 2000, 9 (11): 1972-1977 [11] Han Liang, Tian Fengchun, Wang Yu. Computation of the wavelet and scaling functions using coefficients of wavelet filter. Journal of Chongqing University (Natural and Science Edition), 2007.4: 79-82 Figure 3. Scaling function and wavelet function constructed by the four rational filters Figure 4. Amplitude-frequency characteristic and phase-frequency characteristic of the four rational filters V. CONCLUSION This paper for rational wavelet filter design adopts the method of reducing vanishing moment order, but it also decreases the wavelet orthogonal compactly-supported performance, which is an unpleasant phenomenon in image compression field. But this kind of wavelet filter also has two main advantages: (1) the rational coefficients of wavelet filter can avoid truncation operation when it is operated on computer, and it can completely meet the requirements of perfect-reconstruction condition and vanishing moment condition. It is especially suit for embedded processor applications, for rational wavelet transformation can be achieved by shift operation and multiply-add operation with fast speed and high precision; (2) more information exists in high frequency domain is not favorable for image compression, but in the edge detection and morphological analysis fields, it can provide more information. This paper adopts the thoughts of biorthogonal wavelet filter design, and then the four filters all have linear phase, which is shown in Figure 4. The linear phase of filter, which is just the advantage of biorthogonal wavelet structure, can 80