EFFICIENT IMAGE COMPRESSION ALGORITHMS USING EVOLVED WAVELETS

Transcription

1 International Journal of ISSN Systems and Technologies IJST Vol.4, No.2, pp KLEF 2011 EFFICIENT IMAGE COMPRESSION ALGORITHMS USING EVOLVED WAVELETS 1.G.CHENCHU KRISHNAIAH* 2.T.JAYACHANDRAPRASAD 3.M.N.GIRI PRASAD 1.ECE DEPT.,GKCE,Sullurpet ,A.P,India. 2.ECEDEPT.,RGMCET,Nandyal ,A.P,India 3.ECEDEPT.,JNTUCE,Pulivendula ,A.P,India Abstract: The 9/7 and 5/3 lifting based wavelet filters are widely used in different image compression schemes, such as JPEG 2000 image compression standard. The performance of a hardware implementation of the 9/7 and 5/3 filter banks depends on the accuracy with which filter coefficients are represented. In this paper an attempt has been made to study the performance of 9/7 and 5/3 wavelets on photographic images (monochrome and color) and estimated Peak Signal to Noise Ratio (PSNR), Compression Ratio (), Mean Square Error (MSE), Encoding Time, Decoding Time, Transforming Time or Decomposition Time etc. This study shows that the 5/3 wavelet transform out perform the 9/7 wavelet transform. Key words: Low complexity, efficient, 5/3 wavelet filter, 9/7 wavelet filter, implementation, image compression, Algorithms, Evolved Wavelets. Introduction The Discrete Wavelet Transform (DWT) has gained wide popularity due to its excellent decorrelation property[1], as a consequence many modern image and video compression systems embody the DWT as the transform stage [2], [3]. It is widely recognized that the 9/7 filters [4] are among the best filters for wavelet based image compression [5]. In fact the JPEG2000 image coding standard [6] employs the 9/7 filters as the default wavelet filters for lossy compression, fostering many research efforts in the development of fast and efficient hardware codecs. 127

2 G.Chenchu Krishnaiah The performance of a hardware implementation of the 9/7 filter bank depends on the accuracy with which filter co-efficient are represented. However high precision representation increases hardware resources and processing time. To reduce the complexity of the 9/7 filters the lifting scheme [7] can be adopted. Unfortunately the lifting scheme increases hardware timing accumulation due to its serial nature [8], so that for certain applications it cannot be employed. The flipping structure [8] is an attractive alternative to the standard lifting scheme DWT, since it reduces timing accumulation, however it still requires multiplications. Complexity reduction can be achieved resorting to a filter bank implementation; in particular very good results can be obtained with the cascaded method proposed in [9]. The basic idea described in [9] is to minimize the number of bit required to represent the 9/7 coefficients. Since this operation would move filters zeros from their original position, the authors modify some terms to account for zeros compensation. Currently the compatibility of low complexity 9/7 filters implementation with floating point ones has not been stressed yet. The aim of this paper is to show that great complexity reduction can be achieved analyzing the 9/7 filters directly from their analytical derivation [4]. In particular the proposed solution shows negligible quality loss if employed in the direct DWT, with the floating point 9/7 filter bank in the inverse DWT. 1. THEORETICAL DERIVATION Let s consider the filter bank shown in figure 1, where filters with length kand l respectively, and are the low pass and high pass analysis 128

3 Efficient Image. the low pass and high pass synthesis ones with length and. It is well known that wavelet filter banks ought to satisfy the perfect reconstruction conditions [10]: Imposing the biorthogonality condition together with filters symmetry ( and G(z) = (-z)) we can rewrite the perfect reconstruction conditions as: (z) = H(-z) As shown in [4], writing the non distortion condition 3 on h and trigonometric polynomials, it becomes: in terms of Moreover, together with divisibility of H and respectively by and [4] it leads to: Where R( ) is an odd polynomial in Cos( ) and 2l=k+. The 9/7 filters have been proposed in [4] as a particular case of trigonometric polynomial that satisfy equation 6 with R 0, k=4 and =4 equation 6 becomes =4. When R 0, k=4 and 8 The term cos( / 2) can be split into two equal parts with degree 4. The polynomial in sin ( / 2) can be considered as a third order equation and factorized into two polynomials with degree 2 and 4 respectively in order to obtain: Where r is the real solution of the third order equation x 10 / 20x x 0 (9) 129

4 G.Chenchu Krishnaiah and b. Equation 8 leads to: cos( / 2) H a b a ( ) [ sin( / 2) sin( / 2) ] (10) Developing equations 10 and 11 we can build filters coefficients as shown in table 1 b 3 1 a r where K 1 = 2 8 b, K 2 2, k3, J1 and J2. 2a 8a 32a 2r 8r 2. THE LIFTING SCHEME The conventional structure of two-channel filter banks based on lifting steps is shown in Figure 2. The incoming signal (z-domain) is split into two paths (polyphase transform) containing the values at even or odd sample positions. In order to devise the filter 130

5 Efficient Image. properties, alternating lifting and dual lifting steps are applied, in which samples from one path are filtered by L i (z) and added to a sample of the other path. With respect to the design method to be proposed, however, the illustration with a different flow chart is more helpful. Figure 3 depicts a lifting cascade suitable for representing a 9-tap low-pass and a 7-tap high-pass filter pair. Essentially, it shows the signal flow for processing a signal with eight samples x 0 to x 7. A pair of samples at even positions is weighted by (typically negative) coefficients α and added to the sample in between. The next lifting step combines the results of the summations in pairs using the coefficients β. The third and fourth lifting step act in the same manner using the weights γ and δ. The property of integer-to-integer mapping, which will be essential for lossless compression, is simply imposed 131

6 G.Chenchu Krishnaiah by properly rounding the intermediate values to integer values [12] (not shown in the flow diagram). The arithmetic calculations are (with m= 0, 1, 2,... ) d m = x 2m+1 + α(x 2m +x 2m+2 ) a m =x 2m + β (d m 1 + d m ) d m =d m + γ (a m + a m+1 ) (12) a m =a m + δ(d m-1 + d m ) The result after all lifting steps is an interleaved sequence of low-pass filter output a m (approximation signal) and the high-pass filter output d m (detail signal). Figure 3 also shows the reconstruction of the original signal x n by performing the lifting steps in the reverse order and using the opposite signs a m = a m δ (d m-1 + d m ) d m = d m γ (a m + a m+1 ) x = 2m a m β (d m-1 + d m ) (13) x = 2m+1 d m α (x 2m +x 2m+2 ) Furthermore, the flow diagram simply explains the exception handling at the signal borders. When applied to signals with an odd length, the handling has to be changed slightly. 3. FILTER DESIGN 3.1 Filters with maximum number of vanishing moments The flow diagram, as depicted in Figure 3, allows the derivation of the analysis filters of the corresponding two-channel filter bank simply by considering all paths from the input samples to a particular approximation sample a m or detail sample d m, respectively. For the moment we will disregard the rounding operations. The resulting symmetric 7-tap impulse response of the analysis high-pass filter is h 1 [n] = { αβγ βγ [γ (2αβ + 1) + α (1+γβ)] (2βγ + 1) [γ (2αβ + 1) + α (1+γβ)] βγ αβγ } (14) and the analysis 9-tap low-pass filter reads as h 0 [n] = { αβγδ βγδ 132

7 Efficient Image. {δ [γ (2αβ + 1) + α (1 + γβ)] + αβ (1 + γδ)} [δ (2βγ + 1) + β(1 + γδ)] {Α[δ (2βγ + 1) + β(1 + γδ)] + (1 + 2γδ)+ α [δ (2βγ + 1) + β(1 + γδ)]} (15) [δ (2βγ + 1) + β(1 + γδ)] {δ[γ(2αβ + 1) + α (1 + γβ)] + αβ(1 + γδ)} βγδ α βγδ} The synthesis filters are derived by following all paths from a particular approximation (or detail) sample to the reconstructed signal values x n. In this particular lifting structure, it turns out that they are directly related to the analysis filters by g 0 [n] = ( 1) n +1 h 1 [n] n = 0, 1, 2,... (16) g 1 [n] = ( 1) n h 0 [n]. (17) The frequency response (in z-domain) of a t-tap filter h[n] is t 1 H(z) = h[n]z n. (18) n=0 Since h 0 [n] and g 0 [n] should be real low-pass filters, their magnitude responses at sampling frequency must be equal to zero: G 0 (z) z = 1 = 0 and H 0 (z) z= 1 =0. This leads to following conditions in the spatial domain. 0 = αβγ + βγ + [γ (2αβ + 1) + α (1 + γβ)] +(2βγ + 1) + [γ (2αβ + 1) + α (1 + γβ)] +βγ + αβγ. (19) and 0 = αβγδ βγδ +{δ [γ(2αβ + 1) + α (1 + γβ)] + αβ (1 + γδ)} [δ (2βγ + 1) + β(1 + γδ)] +{α [δ (2βγ + 1) + β(1 + γδ)] + (1 + 2γδ) + α [δ (2βγ + 1) + β(1 + γδ)]} (20) 133

8 G.Chenchu Krishnaiah [δ (2βγ + 1) + β(1 + γδ)] + { δ[γ(2αβ + 1) + α (1 + γβ)] + αβ(1 + γδ)} βγδ + αβγδ The original aim of filter design in [13] was to create low-pass filters with frequency responses that are as flat as possible at sampling frequency by imposing a maximum number of so-called vanishing moments, i.e. multiple zeros at H 0 (z) z=-1 and G 0 (z) z=-1. Multiple vanishing moments at z = -1 can be incorporated by substituting z with n n p (p = 0, 1, 2,... ) in equation (18). The second zero for G 0 (z) (and accordingly for H 1 (z) at z = 1), for example, is included using z = n n leading to the condition 0 = 0 αβγ + 1 βγ + 2 [γ (2αβ + 1) + α (1 + γβ)] +3 (2βγ + 1) + 4 [γ(2αβ + 1) + α (1 + γβ)] +5 βγ + 6αβγ. (21) A different interpretation of this approach is based on the approximation of signal segments by polynomials of increasing order [13]. The condition for the second zero for H 0 (z) and G 1 (z) reads as 0 = 0 αβγδ 1 βγδ +2{δ [γ (2αβ + 1) + α (1 + γβ)] + αβ (1 + γδ) } 3 [δ (2βγ + 1) + β(1 + γδ)] + 4{α [δ (2βγ + 1) + β(1 + γδ)] + (1 + 2γδ) + α [δ (2βγ + 1) + β (1 + γδ)]} 5 [δ (2βγ + 1) + β(1 + γδ)] +6{δ [γ (2αβ + 1) + α (1 + γβ)] + αβ(1 + γδ)} 7 βγδ + 8 αβγδ. (22) The conditions (21) and (21) are, however, not independent from (19) and (20). Two more constraints are necessary for the determination of the four weights α... δ. Choosing z = n n 2 imposes another vanishing moment. The corresponding conditions are 134

9 Efficient Image. 0 = 0 αβγ + 1 βγ + 4 [γ (2αβ + 1) + α (1 + γβ)] +9 (2βγ + 1) + 16 [γ (2αβ + 1) + α (1 + γβ)] +25 βγ + 36 αβγ. (23) 0 = 0 αβγδ 1 βγδ +4{δ [γ(2αβ + 1) + α (1 + γβ)] + αβ (1 + γδ)} 9 [δ (2βγ + 1) + β(1 + γδ)] + 16{α [δ (2βγ + 1) + β(1 + γδ)] + (1 + 2γδ) + α [δ (2βγ + 1) + β (1 + γδ)]} 25 [δ (2βγ + 1) + β(1 + γδ)] +36{δ [γ (2αβ + 1) + α (1 + γβ)] + αβ (1 + γδ)} 49 βγδ + 64 αβγδ. (24) Equations (21) (24) form a system of non-linear equations resulting to the irrational weights α β γ (25) δ Due to the inherent structure of the filter bank, each of the conditions impose double zeros, i.e. each filter shows four vanishing moments in total. The result is exactly the same as derived from the factorisation of a polyphase matrix presented in [14]. 4. DESIGN OF 5/3 WAVELET FILTERS Setting the factors γ and δ equal to zero shortens the work in signal decomposition. The lengths of the impulse responses are reduced to 5 taps for the low-pass filter and 3 taps for the high pass, respectively h [ n] { (1 2 ) } 0 h[ n] { 1 } 1 (26) 135

10 G.Chenchu Krishnaiah The required conditions are in z-domain H 0 (z) z=-1 = 0 = αβ β + (1+2 αβ ) - β + αβ H 1 (z) z=1 = 0=α+1+α (27) Leading to the unique solution of α=-1/2 and β=1/4. This is accordance with the original solution in [11]. 5. SIMULATION RESULTS The proposed low-complexity architecture has been tested on the standard images Lena 256 x 256 and Barbara varying the number of decomposition levels. The proposed architecture has been employed for the direct transform whereas a floating point, standard 9/7and 5/3 implementation has been used for the inverse transform. The values of Peak Signal to Noise Ratio (PSNR), Compression Ratio (), Mean Square Error (MSE), Original image size, Compressed image size, Encoding time, Decoding time and transforming time or decomposition time were obtained from the experimental results, and summarized in tables below. The 5/3 filters have lower computational complexicity than the 9/7 s. However the performance gain of the 9/7 s over the 5/3 s is quite large for JPEG

11 Efficient Image. PERFORMANCE COMPARISIONS OF 9/7 & 5/3 LIFTING BASED INTEGER TO INTEGER WAVELET TRANSFORMS INPUT IMAGE PERFORMANCE 5/3-9/7- Cameraman Rice Bird Encoding Time (sec) Decoding Time (sec) PSNR (db) Encoding Time(sec) Decoding Time(sec) Encoding Time(bits) Decoding Time(bits)

12 ENCODING TIME (SEC) G.Chenchu Krishnaiah SAMPLE GRAPHS: /7transform 5/3 transform TYPE OF WAVELETS Figure 1: Encoding time values of 9/7 &5/3 wavelet transforms for Cameraman image (monochrome) 138

13 M S E (db) Decoding TIME (SEC ) Efficient Image /7transform 5/3 transform TYPE OF WAVELETS Figure 2: Decoding time values of 9/7 &5/3 wavelet transforms for Cameraman image (monochrome) /7transform 5/3 transform 139 TYPE OF WAVELETS

14 C R (bpp) P S N R (db) G.Chenchu Krishnaiah Figure 4: PSNR values of 9/7 &5/3 wavelet transforms for /7transform 5/3 transform TYPE OF WAVELETS /7transform 5/3 transform TYPE OF WAVELETS Cameraman image (monochrome) Figure 5: Compression Ratio () values of 9/7 &5/3 wavelet transforms for Cameraman image (monochrome) 140

15 Efficient Image Dog Compressed image Size(bits) Barbara Compressed image Size(bits) Rose Compressed image Size(bits) Circuit Original Image Size (bits) Compressed image Size(bits)

16 G.Chenchu Krishnaiah Pepper Compressed image Size(bits) Gold Hill Compressed image Size(bits) Lena Compressed image Size(bits) Gray Granite

17 Efficient Image Compressed image Size(bits) Circle Compressed image Size(bits) Hill Compressed image Size(bits) GKCE Font

18 G.Chenchu Krishnaiah Compressed image Size(bits) GKCE Logo Compressed image Size(bits) Bridge Original Image Size(dB) Compressed image Size(dB) CONCLUSION In this paper a low-complexity, efficient 9/7 wavelet filters implementation, has been derived. A detailed analysis of the proposed solution architectural impact has been shown with performance and comparisons with the direct implementation. 144

19 Efficient Image. We have presented a new biorthogonal 9/7 tap wavelet with simple coefficients, so computational complexity is reduced greatly compared to the wellknown CDF 9/7 tapwavelet. The simulation shows that the new 9/7-tap wavelet is very ideal alternative to CDF 9/7 tap wavelet. The other wavelet transform, 5/3 wavelet is very efficient in lossless compression and is low complex. From all the above factors, we can conclude that simple 9/7 and 5/3 wavelet transforms are very efficient than the conventional wavelets/ traditional wavelets/ hand designed wavelets presently used to compress the images. REFERENCES 1. M. Vetterli and J. Kovacevic, Wavelets and Subband Coding, SignalProcessing, Prentice Hall, Englewood, Cliff. NJ, D. Taubman, High performance scalable image compression with EBCOT, IEEE Trans. On Image Processing, vol.9, No.7, pp , Jul A. Said and W.A. PLearlman, a new, fast, and efficient image codec based on Set Partitioning In Hierarchical Trees, IEEE Trans. On Circuits and Systems for Video Technology, Vol.6, no.3, pp , Jun M. Antonini, M. Barlaud, P. Mathieu, and I. Daubechies, Image coading using the wavelet transform, IEEE Tran. On Image Processing, Vo.1, No.2,,pp , Apr J.D. Villasenor, B. Belzer, and J. Liao, wavelet filter evaluation for image compression, IEEE Tran. On Image Processing, Vol.4, No.8, pp , Aug M. Boliek, JPEG 2000 Final Committee Draft, I. Daubechies and W. Sweldens, Factoring Wavelet Transforms into Lifting Steps,: Tech.Rep.Bell Laboratories, Lucent Technologies,

20 G.Chenchu Krishnaiah 8. C.T. Haung, P.C. Tseng, and L.G. Chen, Flipping Structure: an efficient VLSI architecture for liftingbased discrete wavelet transform, IEEE Tran. On Signal Processing, vol.52, no.4, pp , Apr K.A. Kotteri, A.E. Bell, ad J.e. Carletta, Design of multiplierless, highperformance, wavelet filter banks with image compression applications, IEEE Tran. On Circuits and Systems-I, vol.51, no.3, pp , Mar G. Strang and T.Q.Nguyen, Wavelets and Filter Banks, Wellesley, Wellesley- Combridge, MA, Sweldens, W; The lifting scheme: A new philosophy in biorthogonal wavelet construction on proc. Of SPIEE, Vo..2569, Sar Diego, USA, July 1995, Calgerbank, A.R.; Daubechies, I.; Sweledens, W.; Yeo, B.L,: Wavelet Transform that maps integers to integers. Applied Computational and harmonic analysis, Vol.5, No.3, 1998, Strutz, T.: Muller, E.: Wavelet filter design for image compression. IEEE-SP Int. Symposium on Time-Frequency and Time-scale analysis, Paris, June 1996, Cohen, A.; Daubechies, I.; Feauveau, J.-C.: Biorthogonal Bases of compactly supported Wavelets. Comm. On Pure and Applied Mathematics, Vol.45, 1992,

21 Efficient Image. 147