Image Coding Scheme Based on Object Extraction and Hybrid

Usama S. Mohammed et al. / International Journal of Engineering Science and Technology Vol. 2(5), 2010, 1375-1383

where Sn(T) is the significance of the set of coordinates T, and b(i,j) is the coefficient value at coordinate (i,j). There are two passes in the algorithm, the sorting pass and the refinement pass. The sorting pass is performed on the list of insignificant sets (LIS), list of insignificant pixels (LIP) and the list of significant pixels (LSP). The LIP and LSP consist of nodes that contain single pixels while the LIS contains nodes that have descendants. The maximum number of bits required to represent the largest coefficients in the spatial orientation tree is obtained and designed as nmax and is given by

}j)b(i,{max

j)(i,log2nmax (2)

During the sorting pass, those coordinates of the pixels which remain in the LIP are tested for significance by using equation (1). The result Sn(T) is sent to the output. Those that are significant will be transferred to the LSP as well as have their sign bit output. Sets in the LIS will also have their significance tested and if found to be significant, will removed and partitioned into subsets. Subsets with a single coefficient and found to be significant will be added to the LSP, or else they will be added to the LIP. During the refinement pass, the n-th most significant bit of the coefficients in the LSP is output. The value of n is decreased by 1 and the sorting and refinement passes occur again. This continues until either the desired rate is reached or until n=0 and all nodes in the LSP have all their bits output. It is clear from equation (1) and (2) that the coding performance of the SPIHT coding algorithm is highly related to the distribution of the b(i,j) in the 3P+1 subband images, where P is the number of scales. So, the normalization of the filter coefficients is indeed an efficient bit allocation scheme to put more bits in coding of lower frequency subband images. Moreover, to reach the higher coding performance, more scales in the wavelet decomposition are required. The main disadvantage of the SPIHT coding algorithm is that the error in one bit of the coded bit stream will result in a noticeable error in output of the decoder.

III. PROPOSED OBJECT-BASED HYBRID IMAGE CODING (OB-HIC) ALGORITHM In this work, combination of the object-based DCT coding and the high performance of the set partitioning in hierarchical tree (SPIHT) coding is used. The modification of the subband image data in the wavelet domain is done based on the DCT transformation and the classification of the wavelet coefficients in the LL subband. The modified data size is exactly the same size as the original one and containing almost the same original information but having smaller values. The proposed hybrid coding algorithm is described next. First, one-level of the discrete wavelet transform is applied on the input image. This process will generate 4-subband

data (xll, xlh, xhl, xhh) in this layer. Next, the baseband image j)(i,xll is compressed using the object-based DCT

coding method (this process is reported in Section 4), and its reconstructed image is designed as ),(' jixll. The

difference between ),( jixll and ),(' jixll is the residual basedband image ),('' jixll , i.e.

).j,i(x)j,i(x)j,i(x 'llll

''ll = (3)

Weber’s law indicates that for a wide range of image intensity, the ratio of the just-noticeable difference (j.n.d) to the image intensity is a constant. To make use of this phenomenon in the proposed coding algorithm, the subband

image ),('' jixll , ),( jixlh , ),( jixhl and ),( jixhh will be normalized as indicated below:

kjix

jixjix

ll

llnll

),(

),(),(

'

''

kj)(i,x

j)(i,xj)(i,x

'll

lhnlh

kjix

jixjix

ll

hlnhl

),(

),(),(

' kj)(i,x

j)(i,xj)(i,x

'll

hhnhh

(4)

Where k is a constant used to minimize the number of the sorting passes in the SPIHT coder, and will be determined as follows:

2/))(*2)(( '''llll XMaxXMaxK (5)

As in predictive image coding [20], the absolute values of the normalized coefficients will be modified as follows:

ISSN: 0975-5462 1377


qpjiyjiyjiy np

np

np ,...,1,),(2),(),( 11 (6)

Where j)(i,yn0 is the value of the normalized coefficients ),( jixn

ll , ),( jixnlh , ),( jixn

hl and ),( jixnhh , and q is an

integer constant which may change with different subbands. After normalization and mapping, four sub-images are

generated and designated as ),( jiynll , ),( jiyn

lh , ),( jiynhl and ),( jiyn

hh , which are then located in the corresponding

positions of ),( jixll , ),( jixlh , ),( jixhl and ),( jixhh in the wavelet decomposition. Then, the data in ),( jiynll is rearranged

into four sub-images,

(2i,2j)yj)(i,y nll1

nll =

1,2j)(2iyj)(i,y nll2

nll

1)(2i,2jyj)(i,y nll3

nll (7)

1)1,2j(2iyj)(i,y nll4

nll

This rearrangement is further applied on ),(1 jiynll and so on to obtain hierarchical representation of ),( jiyn

ll , ),( jiy nlh ,

),( jiynhl and ),( jiyn

hh to be coded by the SPIHT coder. SPIHT image coding without adaptive multilevel arithmetic

coding is finally applied on the resulting hierarchical representation to generate symbol streams.

IV. OBJECT-BASED DCT IMAGE CODING

In general, the idea behind the object based DCT image coding is base on modification of the quantized image based on pre-processing (object edge extraction). The modification will be done after the quantization step. The image is subdivided into a block of pixels and then these blocks is classified into edge and non-edge blocks. For the non-edge blocks, a certain number of these block will be located in the top left hand corner and multiplied the rest of our DCT coefficients with 0 (mask matrix). This would simplify the coding process and improve the compression ratio, but the quality of the compressed image will be reduced. This “mask” matrix determined what dimension of the upper left-hand corner of quantized DCT coefficients would be kept and the rest of the coefficients multiplied by 0. In this work, the LL band image is segmented into region of interest (ROI), which is considered important, and background, which is less important. By allowing the ROI to be coded with higher fidelity than background, a high compression ratio with good quality in the ROI can be achieved. Therefore, the greatest benefit of ROI coding is its capability of delivering high reconstruction quality over certain spatial regions at high compression ratios. The object-based DCT image coding process is divided into two steps in the wavelet domain. First is the identification (classification) process and second is the compression process. To identify the ROI, the LL image edges are detected by rainfalling watershed technique [18], and then a morphological filter is used to fill in holes and small gaps. In the rainfalling watershed technique, the threshold value is calculated automatically using the maximum cross-entropy methods [21]. After the classification process, all the background-area is compressed using only the DC coefficient of one block from the background and the foreground-area is compressed using all the DCT coefficients. In the rainfalling watershed technique, the threshold, using in the edge detection process, is obtained automatically using the entropy-based methods [21]. It is based mainly on maximizing the cross-entropy between the edge and non-edge levels. Automatically threshold estimation technique can be summarized as follows: Assume that the foreground and background probability mass function (pmf) is expressed as Pf(g), 0≤ g ≤T, and Pb(g), T+1≤ g ≤G, respectively, where G is the maximum gray level, T is the threshold. The foreground and background area probabilities are calculated as follows:

T

gff gpPTP

0

)()(

G

Tgbb gpPTP

1

)()( (8)

then the Shannon entropy parametrically dependent upon the threshold T for the foreground and background is formulated as:

T

gfff gPgPTH

0

)(log)()( , )(log)()(1

gPgPTH b

G

Tgbb

(9)

ISSN: 0975-5462 1378


the sum of these two entropy expressed by H(T)=Hf(T)+Hb(T). The optimum threshold can be viewed as the threshold that makes the summation of the background and foreground entropy be maximum, which can be formulated by:

Topt=arg max [Hf(T)+Hb(T)] (10) Figure 1 shows the result of the object edge detection using the rainfalling watershed technique.

Fig. 1 the Lena image and the result of applying the rainfalling watershed technique on the image

V. SIMULATION RESULTS

The performance of the proposed techniques is introduced in this section. Various kinds of gray images with size of 512×512 pixels and 8 bpp are selected as a test data. The performance is evaluated by the peak signal-to-noise ratio (PSNR) in dB and the perceptual quality of the reconstructed images. The proposed method is compared with JPEG, JPEG2000, EZBC, SPIHT, EZW, SPECK, Yu and Mitra [15], and the HS-HIC [16]. The 9-tap low-pass filter and the 7-tap high-pass filter are used to decompose the input image into four sub-images. The resulting baseband image was coded with the object-base DCT coder followed by adaptive arithmetic code at moderate bit rates. The constant k in the normalization was automatically calculated from equation (5) and the rearrangement in equation (7) was performed to obtain a 6-scale hierarchical data structure. In the simulation results the exact number of bits used is selected as follows:

For Lena image we used 32446 bits in DCT coding, and 322 bits in SPHIT coding so that the total bit rate per pixel (bpp) is 0.125. To achieve total bit rate per pixel (bpp) is 0.25, we used 65214 bits in DCT coding, and 322 bits in SPHIT coding. The total bit rate per pixel of 0.5 bpp is achieved using 130429 bits in DCT coding, and 643 bits in SPHIT coding.

For Barbara image we used 32397 bits in DCT coding, and 371 bits in SPHIT coding so that the total bit rate per pixel (bpp) is 0.125. To achieve total bit rate per pixel (bpp) is 0.25, we used 57551 bits in DCT coding, and 7985 bits in SPHIT coding. The total bit rate per pixel of 0.5 bpp is achieved using 127517 bits in DCT coding, and 3555 bits in SPHIT coding.

For Goldhill image we used 25264 bits in DCT coding, and 7504 bits in SPHIT coding so that the total bit rate per pixel (bpp) is 0.125. To achieve total bit rate per pixel (bpp) is 0.25, we used 65214 bits in DCT coding, and 322 bits in SPHIT coding. The total bit rate per pixel of 0.5 bpp is achieved using 130429 bits in DCT coding, and 643 bits in SPHIT coding.

Table 1 shows the resulting PSNR values at different bit rates and the reconstructed images using different image coders are shown in Fig.2 and Fig.3. The results clearly show that the OB-HIC coding performance is better than the other image coding performance (about 0.15- 4dB) with the current images used.

ISSN: 0975-5462 1379


Table 1: Comparative analysis of coding efficiency

Image Lena Barbara Goldhill No. of background block= 64 66 0 No. of foreground block= 960 958 1024

CR=16 (bpp=0.5)

HS-HIC [16] 38.262 32.383 33.439 Yu & Mitra [15] 38.2802 33.5239 34.4228

JPEG 35.88 31.73 32.38 JPEG 2000 37.27 32.87 33.24

EZBC 37.47 32.15 33.47 SPECK 37.1 31.54 33.03 EZW 36.28 30.53 32.87

SPIHT 37.25 32.10 33.13 OB-HIC 38.413 33.7046 35.646

CR=32 (bpp=0.25)

HS-HIC [16] 35.092 28.129 30.57 Yu & Mitra [15] 35.0035 31.5928 32.9558

JPEG 32.49 27.77 29.78 JPEG 2000 34.15 28.89 30.53

EZBC 34.35 28.25 30.74 SPECK 34.03 27.76 30.5 EZW 33.17 26.77 30.31

SPIHT 34.14 28.13 30.56 OB-HIC 35.9591 32.8148 33.898

CR=64 (bpp=0.125)

HS-HIC [16] 33.86 25.16 28.98 Yu & Mitra [15] 32.2498 29.0300 29.5140

JPEG 28.9 24.1 26.9 JPEG 2000 30.71 25.24 28.31

EZBC 31.33 25.33 29.09 SPECK 31.02 24.93 28.39 EZW 30.23 24.03 28.24

SPIHT 31.10 24.86 28.48 OB-HIC 34.0140 31.6355 31.913

(a) OB-HIC at 0.125 bpp.

(b) Yu & Mitra [15] at 0.125 bpp.

ISSN: 0975-5462 1380


(c) SPIHT at 0.125 bpp.

Fig. 2 The results of SPIHT, Yu and Mitra [15], and OB-HIC image coder at 0.125 bpp bit rate.

(a) OB-HIC at 0.5 bpp.

(b) Yu & Mitra [15] at 0.5 bpp.

Fig. 3 The results of Yu and Mitra [15] and OB-HIC image coder at 0.5 bpp bit rate. To demonstrate the effect of the hybrid approach in OB-HIC technique, the numbers of significant coefficients were calculated in all subbands in the wavelet domain during the coding process to compare the results of the proposed technique related to the results of the SPIHT coder and the HS-HIC coder in each subband. Only 3-layers are used in this test. Table 2 provides the number of significant coefficients in each subband for the OB-HIC, HS-HIC and SPIHT of Lena image for five passes of the coding process. The bit rate results in this table are calculated without arithmetic coding. The results in this table reveal the performance gain in each subband of the proposed coder (OB-HIC) related to the other coders. As expected, the number of significant coefficients in each pass of the OB-HIC coder is lower than the number of significant coefficients in each pass of the SPIHT coder and the HS-HIC coder. Simulation results show that, in most cases, the binary non-coded bit rate of OB-HIC coder is much better than that of SPIHT and HS-HIC coder. Table 3 and Table 4 provide the same introduced comparison in Table 2 for Barbara image (4-pass) and Goldhill image (5-pass), respectively.

ISSN: 0975-5462 1381


Table 2: List of significant coefficients comparison between OB-HIC, HS-HIC and SPIHT coding performance at each pass of the coding process for Lena image

Subband Lena OB-HIC (PSNR1 = 35.7713dB) , bitrate=0.5 bpp

SubbandHS-HIC (PSNR=33. 9041 dB),

Bitrate =0.4366 bpp Subband

Lena SPIHT (PSNR=31.4271 dB), bitrate=0.3764 bpp

LL3 10 148 51 36 29 LL3 853 1235 1049 398 243 LL3 853 1235 1049 398 243 LH3 0 0 0 8 37 LH3 0 19 141 363 583 LH3 0 18 138 266 457 HL3 0 0 2 8 28 HL3 6 81 305 561 612 HL3 2 83 301 502 580 HH3 0 0 0 1 33 HH3 0 5 61 229 467 HH3 0 1 56 201 365 LH2 0 0 61 404 1039 LH2 0 6 137 699 1435 LH2 0 0 38 294 885 HL2 0 0 6 5 48 HL2 0 11 165 707 1477 HL2 0 0 137 699 1435 HH2 0 0 6 152 574 HH2 0 0 8 127 505 HH2 0 0 1 87 520 LH1 0 0 0 55 370 LH1 0 0 0 34 640 LH1 0 0 0 26 567 HL1 0 0 7 304 1639 HL1 0 0 6 280 1469 HL1 0 0 0 214 1371 HH1 0 0 0 0 35 HH1 0 0 0 1 87 HH1 0 0 0 0 24 Total 10 148 133 973 3832 Total 859 1357 1872 3399 7518 Total 855 1343 1720 2687 6397

Table 3: List of significant coefficients comparison between OB-HIC, HS-HIC and SPIHT coding performance at each pass of the coding process for Barbara image

Subband barbara OB-HIC (PSNR1=32.7856 dB) ,

bitrate=0.25bpp Subband

HS-HIC (PSNR= 28.1294 dB),

Bitrate = 0.33 bpp Subband

SPIHT (PSNR=25.097dB), Bitrate = 0.2455 bpp

LL3 10 82 82 53 LL3 856 1369 957 545 LL3 856 1369 957 545 LH3 0 1 10 24 LH3 0 25 141 432 LH3 0 25 136 362 HL3 0 0 6 19 HL3 0 56 271 517 HL3 0 58 257 534 HH3 0 0 0 17 HH3 0 0 23 212 HH3 0 0 6 168 LH2 0 0 5 43 LH2 0 0 148 1027 LH2 0 0 53 308 HL2 0 0 0 22 HL2 0 9 207 941 HL2 0 0 150 1038 HH2 0 0 0 5 HH2 0 0 47 448 HH2 0 0 0 386 LH1 0 0 10 54 LH1 0 0 0 2 LH1 0 0 0 0 HL1 0 12 1160 4931 HL1 0 0 418 2733 HL1 0 0 30 2154 HH1 0 0 0 3 HH1 0 0 0 346 HH1 0 0 0 1 Total 10 95 1273 5171 Total 856 1459 2212 7203 Total 856 1452 1589 5495

Table 4: List of significant coefficients comparison between OB-HIC, HS-HIC and SPIHT coding performance at each pass of the coding process for Goldhill image

Subband Goldhill OB-HIC (PSNR1 =33.8981

dB) , bitrate=0.25bpp Subband

HS-HIC (PSNR=30.1 dB), Bitrate = 0.4252 bpp

SubbandSPIHT (PSNR= 29.9333 dB),

Bitrate = 0.4096 bpp LL3 18 0 1 135 1069 LL3 641 1207 884 611 392 LL3 641 1207 884 611 392 LH3 0 0 0 190 1009 LH3 0 14 187 577 980 LH3 0 12 177 577 918 HL3 0 0 0 200 1073 HL3 2 27 190 414 685 HL3 1 34 206 373 610 HH3 0 0 1 203 1047 HH3 0 0 20 171 565 HH3 0 0 11 136 445 LH2 0 0 2 457 3149 LH2 0 0 75 550 1208 LH2 0 0 27 491 1574 HL2 0 0 1 488 3521 HL2 0 1 66 514 1328 HL2 0 0 75 550 1208 HH2 0 0 3 234 2378 HH2 0 0 2 49 456 HH2 0 0 0 20 354 LH1 0 49 296 1643 6213 LH1 0 0 0 65 1173 LH1 0 0 0 40 1138 HL1 0 14 416 2133 6337 HL1 0 0 1 148 1351 HL1 0 0 0 166 1371 HH1 0 0 0 20 1311 HH1 0 0 0 0 85 HH1 0 0 0 0 15 Total 18 63 720 5703 27107 Total 643 1249 1425 3099 8223 Total 642 1253 1380 2964 8025

VI. CONCLUDING REMARKS

In this paper, object-based hybrid image coding algorithm is introduced. The proposed algorithm works much more efficiency than the SPIHT coding method. The new distribution of the wavelet coefficients provides reduction in computational complexity. For low bit rate image coding only one sorting pass in the wavelet domain is applied. The simulation results indicate that the PSNR performance of the proposed algorithm is much higher than that of SPIHT, EZW, SPECK, EZBC, and the JPEG-2000 test coder. The advantage in performance of OB-HIC lays in the DWT maps with small coefficients in all subband image data.

ISSN: 0975-5462 1382


References [1] Usama Sayed, Image Coding Technique Based on Object-Feature Extraction, In the Proceedings of the National Radio Science Conference

(NRSC'2005), Cairo, Egypt, CD (Commission C16), (2005). [2] K. An, M Lee, J. Shin, Saliency map model based on the edge images of natural scenes, International Joint Conference on Neural Networks

(IJCNN), USA, (2002). [3] B. Ko, S Kwak, H. Byun, SVM-based salient region(s) extraction method for image retrieval, 17th International Conference on Pattern

Recognition (ICPR 2004), Cambridge, UK, 977-980 (2004). [4] G.P. Nguyen, M. Worring, An user based framework for salient detail extraction, IEEE International Conference on Multimedia and Expo

(ICME), Taipei, Taiwan, (2004). [5] J. Shapiro, Embedded image coding using zerotress of wavelet coefficients, IEEE Trans. Signal Processing, 41:3445-3462 (1993). [6] O. Egger, A. Nicoulin, W. Li, Embedded zerotree based image coding with low decoding complexity using linear and morphological filter

banks, in: Proceeding of IEEE International Conference on Acoustics, Speech and Signal Processing, New York, NY, USA, 2237-2240 (1995).

[7] A. Said and W. Pearlman, A new, fast and efficient image codec based on set partitioning in hierarchical trees, In IEEE Trans. Circuits and Systems for Video Technology, 3: 243-250 (1996).

[8] D. Taubman, High performance scalable image compression with EBCOT, IEEE Trans. Image Process, 7: 1158-1170 (2000). [9] ISO/IEC JTC1/SC29/WG1 N871 R, Embedded, Independent block-based coding of subband data, July 1998. [10] ISO/IEC JTC1/SC29/WG1 N1020 R, EBCOT: Embedded block coding with optimized truncation, October 1998. [11] A. Said and W. Pearlman, Low-complexity waveform coding via alphabet and sample-set partitioning, In Visual Communications and Image

Processing '97, Proc. SPIE 3024, 25-37 (1997). [12] C. Chrysafis, A. Said, A. Drukarev, A. Islam, W. A. Pearlman, SBHP-a low complexity wavelet coder, in IEEE International Conference on

Acoustics, Speech and Signal Processing (ICASSP 2000), Istanbul, Turkey, (2000). [13] W. A. Pearlman, A. Islam, N. Nagaraj, A. Said, Efficient, low-complexity image coding with a set-partitioning embedded block coder, IEEE

Trans. Circuits Syst. Video Technol, 11: 1219-1235 (2004). [14] Shih-Ta Hsiang, J.W.Woods, Embedded image coding using zeroblocks of subband/wavelet coefficients and context modeling, MPEG-4

Workshop and Exhibition at ISCAS 2000, Geneva, Switzerland, (2000). [15] T. Yu and S. K. Mitra, Wavelet based hybrid image coding scheme, IEEE international symposium on circuits and systems, 1: 377-380

(1997). [16] Usama Sayed, Highly Scalable Hybrid Image Coding Scheme, In the Digital Signal Processing Journal, Science Direct, 18(3): 364-374

(2008). [17] S. Singh, V. Kumar, and H. K. Verma, DWT-DCT hybrid scheme for medical image compression, Journal of Medical Engineering &

Technology, 31(2): 109–122 (2007). [18] Wei, Haiping; Zhao, Baojun; He, Peikun, Hyperspectral image compression using SPIHT based on DCT and DWT, Proceedings of the

SPIE, 6787: 67870H (2007). [19] M. Antonini, M. Barlaud, P. Mathieu, and I. Daubechies, Image coding using wavelet transform, IEEE Trans. Image Processing, 1: 205-220

(1992). [20] T. Yu and S. K. Mitra, a novel DPCM algorithm using a nonlinear operator, Proc. IEEE international conference on image processing'94,

Austin, Texas, USA, 871-875 (1994). [21] B. Sankur and M. Sezgin, Image thresholding techniques: A survey over categories, Pattern Recognition, (2001).

ISSN: 0975-5462 1383

Documents

Image Coding Scheme Based on Object Extraction and Hybrid