1. 1 Photographic Tone Reproduction for Digital Images Brandon Lloyd COMP238 October 2002

2. 2 Problems with High Dynamic Range (HDR) The range of light in the real world spans 10 orders of magnitude! A single scene’s luminance values may have as much as 4 orders of magnitude difference A typical CRT can only display 2 orders of magnitude Tone-mapping is the process of producing a good image of HDR data

3. 3 Zone System Used by Ansel Adams. Utilizes measured luminance to produce a good final print Zone: an approximate luminance level. There are 11 print zones Middle-grey: Subjective middle brightness region of the scene, typically map to zone V Key: Subjective lightness or darkness of a scene

4. 4 Zone System Measure the luminance on a surface perceived as middle-gray - map to zone V Measure dynamic range from both light and dark areas. If dynamic range < 9 zones then full range can be captured in print Otherwise dodging-and-burning must be used to bring out details Dodging-and-burning: Witholding or adding light in development to lighten or darken the final print

5. 5 Algorithm First apply a scaling to the whole image. This similar to setting the exposure for mapping to middle-gray Apply automatic dodging-and-burning to compress dynamic range if necessary

6. 6 Luminance Scaling Use log-average luminance to approximate the key of the scene In a normal-key image middle-gray maps to a key value a =.18 suggesting the function:

7. 7 Luminance Scaling

8. 8 Control burn out of high luminances

9. 9 Automatic Dodging-and-Burning Think of this as local adaptation, choosing a key value for every pixel Need a properly chosen neighborhood Dodging-and-burning is applied to regions bounded by large contrasts Use center-surround functions to measure local contrast at different scales

10. 10 Automatic Dodging-and-Burning The effects of using different scales Center Surround s1s1 s2s2 s3s3 s1s1 s2s2 s3s3

11. 11 Automatic Dodging-and-Burning Use difference of Gaussians for center- surround function

12. 12 Automatic Dodging-and-Burning Choose largest neighborhood around a pixel with fairly even luminances Take the largest scale that doesn’t exceed a contrast threshold: Final local operator

13. 13 Automatic Dodging-and-Burning Details recovered by using dodging-and-burning

14. 14 Results Used FFT to compute the Gaussians 8 discrete scales ranging from 1 pixel to 43 increasing by a factor of 1.6  = 0.05

15. 15 Results

16. 16 Results

17. 17 Results

18. 18 Comparison Durand et al. Reinhard et al.

19. 19 Comparison Durand et al. Reinhard et al.