High Dynamic Range Reduction
via Maximization of Image Information

A. Ardeshir Goshtasby

CSE Dept., Wright State University


Abstract: An algorithm has been developed that blends multiple exposure images to produce an image with maximum information content. The algorithm first partitions the image domain into subimages and for each subimage identifies the image that contains the most information within that subimage.The selected images are then blended together using monotonically decreasing blending functions that have a sum of 1 everywhere in the image domain. Chroma is used to measure image information and subimage size as well as width of the blending functions are determined in a gradient ascent algorithm to maximize information content in the created image. An image obtained by this algorithm faithfully reproduces color and contrast in local areas. To further enhance the obtained image, traditional image enhancement methods are applied.


Fig. 1. (a) A color image with a varying luminance. (b) The same image with a constant luminance.
(See the images in high resolution)

Fig. 2. (a)--(e) Multiple exposure images taken of an office room. (f) Image obtained by composing the 12 subimages (image blocks) selected from the five images. (g)-(r) Images corresponding to the selected blocks when multiplied by the blending functions. (s) Image obtained by blending (adding together) images (g)-(r). (t) The most informative image obtained from images (a)-(e).
(See the images in high resolution)
(Download high-resolution images 2a, 2b, 2c, 2d, and 2d)

Fig. 3. (a) Logarithmic mapping of Fig. 2t. (b) Image obtained by blending images 2t and (a) to maximize information content. (c) Image obtained by deblurring image (b) with filters [b a b] and [b a b]^t where a=0.96, b=0.09. (d) Same as in (c) but using a=0.48, b=0.26. (e) Our final result when a=0.56, b=0.22.
(See the imags in high resolution)
(Download high-resolution Fig. 3e)

Fig. 4. Top row shows fifteen original images. The images in the bottom row from left to right show results obtained by Tumblin and Turk [1], Fattal et al. [2], and us. Thanks to Paul Debevec for the original images.
(See the images in high resolution)

FIg. 5. Top row are nine images representing different exposures of a Belgium house, and the three images below that from top to bottom show results obtained by the methods of Tumblin and Turk [1], Fattal et al. [2], and that described here. Thanks to Dani Lischinski for the original images.
(See the images in high resolution)

Fig. 6. Top row shows eleven original images, and images below that from left to right show results obtained by Tumblin and Turk [1], Fattal et al. [2], and us. Thanks to Jack Tumblin for the original images.
(See the images in high resolution.)

Fig. 7. Top row shows four images obtained at different exposure levels of a store window, the bottom-left image shows the result obtained by the method of Fattal et al.[2], and the bottom-right image shows the image obtained by the proposed method. Thanks to Shree Nayar for the original images.

Fig. 8. Top row five images represent the original images of chairs, and the image in the bottom-left shows the result obtained by the method of Fattal et al. [2], and the image in the bottom-right shows the result obtained by our method. Thanks to Shree Nayar for these images also.

Fig. 9. (left) Two images of Scottish Highlands obtained at different exposure levels. (middle) Result obtained by photographic techniques by Max Lyons. (Right) Result obtained by our method. Thanks to Max Lyons for the original images.
(See the result in medium or high resolution.)

To find out the details of the method, read this paper. The paper uses all three rgb color components rather than the chroma shown above. With all three color components, results are slightly better as anticipated.

High-dynamic range reduction can actually be considered an image fusion process. This is discussed in the paper by the same author published at Image and Vision Computing, vol. 23, 2005, pp. 611-618.

References

[1] J. Tumblin and G. Turk, LCIS: A Boundary Hierarchy for Detail-Preserving Contrast Reduction, Proceedings of SIGGRAPH 1999, 83-90.

[2] R. Fattal and D. Lischinski and M. Werman, Gradient Domain High Dynamic Range Compression, Proceedings of ACM SIGGRAPH 2002, 249-256.


[WSU Home Page] [CSE Department Home Page] [Intelligent Systems Laboratory]

For more information contact A. Goshtasby (agoshtas@cs.wright.edu).

Last modified: 7/28/03.