17 1 2012 1 Journal of Image and Graphics Vol. 17No. 1 Jan. 2012 TP391. 4 A 1006-8961201201-0033-07. NSS J. 2012 171 33-39 NSS 214122 NSS No-reference image quality assessment based on natural scene statistics and wavelet Jin Bo Li Chaofeng Wu Xiaojun School of Internet of Things Engineering Jiangnan University Wuxi 214122 China AbstractTo estimate a range of image distortions a novel no-reference image quality assessment method is proposed based on wavelet multi-scale transformation. For natural scene statistics NSS model the sub-band energy of wavelet transformation has a linear distribution with scale index. According to this principle the energy distribution of ideal image could be predicted from high-scale sub-band energy which was not badly affected by distortion. Meanwhilean effective method for identifying and compensating for an inappropriate distortion was presented. Finally the quality metric was constructed by quantifying the difference between predicted energy and real energy in degradation image. Experimental results showed that the new method was consistent with subjective assessment and outperformed the other methods. Key wordsno-reference image quality assessmentnatural scene statistics NSSmodelwavelet transformmulti-scale predictionenergy compensation 0 3 2011-01-17 2011-04-26 61170120 B12011132 1987 E-mailwxjinbo@ 163. com
34 www. cjig. cn 17 3 1 10 wavelet transform 1 2 JPEG 3 1 2 1 2 4 2 LIVE 11 29 5 6 3 E s o = 1 T Σ log 2 C s o + 1 E T C s o Lu Wen contourlet C 0 7 NSS HVS 0. 1 contourlet 8 Moorthy 9 1 2 3 7 8 JPEG2000 3 9 JPEG2000 JPEG FastFading 1 2 1. 1 8
1 NSS 35 1. 2 bior4. 4 4 JPEG2000 FastFading 3 E 3 E LIVE 29 4 a d e 2 E 4c JPEG 4b DCT 4 LIVE JPEG2000 JPEG FastFading 5 3 Fig. 3 Sub-band energy distribution of ideal image with visual important region 20 Fig. 4 4 Sub-band energy distribution of different distortions with visual important region
36 www. cjig. cn 17 2 3 4 E 1 nn 10 1 I H H s = I s I - 1 4 2 H s s s 3 2 1 3 2 1 I s s I 4 4 2 5 D 4 Fig. 5 Adjust the predicted sub-band energy D 4 2 H s P P P s = D 4 H s 3 1 2 4 3 4d 4 D 4 D s = D s + δt 4 5 5a δ 0 δ < 1 0. P t 4 D 4 IF u 4 U 4 THEN P = I avg 4 5JPEG 4b u 4 D 4 U 4 I avg JPEG 1n 4 I 4 JPEG I 5b JPEG 8 8 3 4 4 4 1 4 2 3 4 c 1 4 2 3 JPEG J c D
1 NSS 37 B outer h = 1?N/4-1 { MN Σ ΣM Xi 1 4+ 4j B inner h = 1?N/4-1 MN Σ ΣM Xi 2 3+ 4j B outer v = 1?M/4 MN Σ -1 { Σ N X 1 j B inner v = 1?M/4 MN Σ -1 Σ N X 2 j 4+ 4i 3+ 4i W s = 2. 60. 192 + 0. 114f s exp- 0. 114f s 1. 1 6 7 9 f s 1 2 P D 8 9 Q Q = Σ 2 f s = f 2 sx + f 2 sy 1 2 fn 10 JPEG2000 JPEG Rayleigh f sx f sy s 5 f n 0. 05 DMOS 7 4 s = 1 W s log 2 1 + P s - D s w+ J c 11 w 1. 2 0. 8 3 LIVE J c = Bi h + B i v - 1 8 11 29 5 B o h + B o v JPEG2000 169 B outer C inner JPEG 175 145 Xi j 1 M N 145 Rayleigh 145 779 J c JPEG 6 DMOS 1 HVS 12 CSF CC 2 Spearman SROCC 1 7-9 Table 1 1 Performance comparison of different methods JP2K JPEG WN BLUR FF ALL 7 0. 852 7 0. 581 0 0. 957 6 0. 891 7 0. 852 7 N /A CC 8 0. 906 5 N /A 0. 969 4 0. 945 4 N /A N /A 9 0. 808 6 0. 901 1 0. 953 8 0. 829 3 0. 732 8 0. 820 5 0. 909 9 0. 898 2 0. 968 3 0. 908 1 0. 851 9 0. 853 6 7 0. 823 8 0. 562 3 0. 600 5 0. 856 1 0. 823 1 N /A SROCC 8 0. 898 1 N /A 0. 950 2 0. 934 7 N /A N /A 9 0. 799 5 0. 891 4 0. 951 0 0. 846 3 0. 706 7 0. 819 5 0. 885 3 0. 866 4 0. 957 3 0. 889 5 0. 838 6 0. 849 6
38 www. cjig. cn 17 6 DMOS 6 Fig. 6 DMOS Scatter of model prediction and DMOS 4
1 NSS 39 References 1Crete F Dolmiere T Ladret P et al. The blur effectperception and estimation with a new no-reference perceptual blur metric C/ / Human Visual and Electronic Imaging Ⅻ. San Jose USA Proceeding of SPIE 200764920I. 2Marziliano PDufaux FWinkler Set al. A no-reference perceptual blur metricc/ / Proceeding of IEEE International Conference on Image Processing. Genimedia SASwitzerland IEEE Computer Society 200257-60. 3Wang Z Sheik H R Bovik A C. No-reference perceptual quality assessment of jpeg compressed imagesc/ / Proceeding of IEEE International Conference on Image Processing. New York USA IEEE Computer Society 2002447-480. 4Pan F Lin X Rahardja S et al. Using edge direction information for measuring blocking artifacts of imagesj. Multidimensional Systems and Signal Processing 2007 184 297-308. 5Charrier C Lebrun G Lezoray O. A machine learning-based color image quality metric C/ /Proceedings of Third European Conference on Color in Graphics. LeedsUKImaging And Vision 2006251-256. 6Babu R V Perkis A. An hvs-base no-reference perceptual quality assessment of jpeg coded images using neural networksc/ / Proceedings of IEEE International Conference on Image Processing. Genoa ItalyIEEE Computer Society 2005433-439. 7Lu WZeng KTao D Cet al. No-reference image quality assessment in contourlet domain J. Neurocomputing 2010 734-6 784-794. 8Lou B Shen H B Zhao W F et al. No-reference image quality assessment based on statistical model of natural imagej. Journal of Zhejiang University Engineering Science 2010 442 248-252.. J. 2010 442 248-252. 9Moorthy K A Bovik A C. A two-step framework for constructing blind image quality indicesj. IEEE Signal Processing Letters 2010 175 513-516. 10Schaaf AHateren J H. Modeling the power spectra of natural imagesstatistics and informationj. Vision Research 1996 36172759-2770. 11Sheikh H RWang Z. Live Image Quality Assessment Database EB /OL 2005 2010-06-14. http/ / live. ece. utexas. edu / research / quality. 12Wang T S Gao X B Lu W et al. A new method for reducedreference image quality assessment J. Journal of Xidian University 2008 351 101-104.. J. 2008 351 101-104.