HIGH DYNAMIC RANGE IMAGE TONE MAPPING BY OPTIMIZING TONE MAPPED IMAGE QUALITY INDEX

Transcription

1 HIGH DYNAMIC RANGE IMAGE TONE MAPPING BY OPTIMIZING TONE MAPPED IMAGE QUALITY INDEX Kede Ma, Hojatollah Yeganeh, Kai Zeng and Zhou Wang Department of ECE, University of Waterloo July 17, 2014

2 2 / 32 Outline 1 2 Experimental Results 3

3 3 / 32 Outline 1 2 Experimental Results 3

4 4 / 32 The Advantage of HDR Images Represent higher precision of luminance levels in natural scenes.

5 5 / 32 The Advantage of HDR Images Represent higher precision of luminance levels in natural scenes. The Acquisition of HDR Images Directly acquired with HDR imaging devices; Indirectly fused from multiple Low Dynamic Range (LDR) images captured at different exposure levels.

6 6 / 32 Objective Compress the dynamic range of HDR images for display on LDR devices.

7 7 / 32 Objective Compress the dynamic range of HDR images for display on LDR devices. Why TMOs? A surrogate for HDR display technology; Show HDR images on printed media; An interesting problem by itself.

8 8 / 32 The Workflow HDR Imaging Workflow HDR reconstruction HDR Images Tone mapping

9 9 / 32 Existing TMOs Literature Review Linear, Gamma and Log-normal mappings; Stevens s law-based algorithm [Tumblin, 1993]; Histogram adjustment technique [Larson, 1997]; Gradient-based algorithm [Fattal, 2002]; Photographic tone mapping [Reinhard, 2002]; Adaptive logarithmic mapping [Drago, 2003]; Display adaptive tone mapping [Mantiuk, 2008]; Linear windowed tone mapping [Shan, 2010]; Multiscale decomposition based algorithm [Gu, 2013].

10 mm 10 / 32 A Visual Example

11 mm 11 / 32 A Visual Example Which algorithm is the best?

12 mm 12 / 32 A Visual Example Which algorithm is the best? Any room for further improvement? If any, how?

13 13 / 32 Subjective Quality Assessment Expensive; Time consuming; Difficult for automatic optimization;

14 14 / 32 Subjective Quality Assessment Expensive; Time consuming; Difficult for automatic optimization; Objective Quality Assessment Traditional quality measures do not apply. Tone mapped image quality index (TMQI).

15 15 / 32 TMQI TMQI Computation in [Yeganeh, 2013] Two key factors: structural fidelity and statistical naturalness. TMQI = as(x, Y) α + (1 a)n(y) β. (1)

16 16 / 32 TMQI TMQI Computation in [Yeganeh, 2013] Two key factors: structural fidelity and statistical naturalness. TMQI = as(x, Y) α + (1 a)n(y) β. (1) Structural fidelity: S local (x, y) = 2 σ x σ y + C 1 σ xy + C 2 σ x 2 + σ y 2, (2) + C 1 σ x σ y + C 2 Statistical naturalness: N = 1 K P m P d. (3)

17 17 / 32 Outline Experimental Results 1 2 Experimental Results 3

18 18 / 32 Experimental Results Mathematical Formulation Structural Fidelity Update Y opt = arg max TMQI(X, Y). (4) Y Ŷ k = Y k + λ Y S(X, Y) Y=Yk. (5) Statistical Naturalness Update (3/L)aŷ y i k+1 = k 0 ŷ i k L/3 (3/L)(b a)ŷ i k + (2a b) L/3 < ŷi k 2L/3 (3/L)(L b)ŷ i k + (3b 2L) 2L/3 < ŷi k L. (6)

19 19 / 32 Structural Fidelity Update Only Experimental Results Tone-mapped Desk Images and Their Structural Fidelity Maps (a) initial image (b) after 10 iterations (c) after 50 iterations (d) after 100 iterations (e) after 200 iterations (f) initial image, S = (g) 10 iterations, S = (h) 50 iterations, S = (i) 100 iterations, S = (j) 200 iterations, S = 0.966

20 20 / 32 Experimental Results Statistical Naturalness Update Only Tone-mapped Building Images (a) initial image, N = (b) 10 iterations, N = (c) 50 iterations, N = (d) 100 iterations, N = (e) 200 iterations, N = 0.971

21 Experimental Results 21 / 32 TMQI comparison between initial and converged images Image Gamma Reinhard s Drago s Log-normal Bridge initial image converged image Lamp initial image converged image Memorial initial image converged image Woman initial image converged image

22 22 / 32 Experimental Results Full version of the proposed algorithm Tone-mapped Images Created by Gamma Mapping (a) S=0.148, N=0.000 (b) S=0.959, N=0.997 (c) S=0.521, N=0.038 (d) S=0.976, N=0.999

23 23 / 32 Experimental Results Full version of the proposed algorithm Cont.1 Tone-mapped Bridge Images Created by [Reinhard, 2002] (a) S=0.847, N=0.764 (b) S=0.971, N=1.000

24 24 / 32 Experimental Results Full version of the proposed algorithm Cont.2 Tone-mapped Lamp Images Created by [Drago, 2003] (a) S=0.787, N=0.966 (b) S=0.970, N=0.999

25 25 / 32 Experimental Results Evolution of Tone-mapped Image Woods with Initial Images Created by Different TMOs. Structural fidelity S Number of iterations Gamma mapping Lognormal mapping Drago s TMO Reinhard s TMO Statistical naturalness N Number of iterations Gamma mapping Lognormal mapping Drago s TMO Reinhard s TMO

26 26 / 32 Experimental Results Evolution of Tone-mapped Image Woods with Initial Images Created by Different TMOs. Structural fidelity S Number of iterations Gamma mapping Lognormal mapping Drago s TMO Reinhard s TMO Statistical naturalness N Number of iterations Gamma mapping Lognormal mapping Drago s TMO Reinhard s TMO Increase monotonically;

27 27 / 32 Experimental Results Evolution of Tone-mapped Image Woods with Initial Images Created by Different TMOs. Structural fidelity S Number of iterations Gamma mapping Lognormal mapping Drago s TMO Reinhard s TMO Statistical naturalness N Number of iterations Gamma mapping Lognormal mapping Drago s TMO Reinhard s TMO Increase monotonically; Converge regardless of initial images;

28 28 / 32 Experimental Results Evolution of Tone-mapped Image Woods with Initial Images Created by Different TMOs. Structural fidelity S Number of iterations Gamma mapping Lognormal mapping Drago s TMO Reinhard s TMO Statistical naturalness N Number of iterations Gamma mapping Lognormal mapping Drago s TMO Reinhard s TMO Increase monotonically; Converge regardless of initial images; Trapped in local optima.

29 29 / 32 Outline 1 2 Experimental Results 3

30 30 / 32 Contribution Proposed an iterative tone mapping algorithm by optimizing TMQI; Future work Develop better quality measures; Classify and quantify distortions introduced by TMOs; Improve the optimization process.

31 31 / 32 References H. Yeganeh and Z. Wang, Objective quality assessment of tone-mapped images, IEEE TIP, J. Tumblin and H. Rushmeier, Tone reproduction for realistic images, IEEE Computer Graphics and Applications, G. W. Larson et al., A visibility matching tone reproduction operator for high dynamic range scenes, IEEE TVCG, R. Fattal et al., Gradient domain high dynamic range compression, ACM TOG, E. Reinhard et al., Photographic tone reproduction for digital images, ACM TOG, F. Drago et al., Adaptive logarithmic mapping for displaying high contrast scenes, Computer Graphics Forum, R. Mantiuk et al., Display adaptive tone mapping, ACM TOG, Q. Shan et al., Globally optimized linear windowed tone mapping, IEEE TVCG, B. Gu et al., Local edge-preserving multiscale decomposition for high dynamic range image tone mapping, IEEE TIP, 2013.

32 32 / 32 Thank you