International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
a new view in the unknown illumination environment. [Shen, 
2000; Wong, 1997a; Wong, 1997b] 
However, for the applications in photogrammetry and remote 
sensing, such method works less significance, because the 
dense sampling in multi angles for satellite and aeroplane is too 
expensive in cost of fund and algorithm. So far, it's not mature 
method for remote sensing and aero-borne images. But for close 
range images, it may be a promising method although the 
algorithm current is not practicable. 
3.3 Image Based Rendering 
The method based on Image Based Rendering (IBR) is a 
compromise solution between the two methods above [Shen, 
2000; Zongker, 1999; Chuang, 2000]. Such method need 
neither restore the model information of the scene completely 
nor interpolate densely. It can be looked as a warp function of 
plenoptic function because it builds the transfer relation 
between known plenoptic function and unknown plenoptic 
function when the scene illumination is changed. The basic idea 
is developed from Environment Mapping [Blinn, 1976] to 
Environment Matting [Zongker, 1999; Chuang. 2000]. The idea 
of Environment Matting takes into account of only the 
reflection and refraction characteristic of each pixel in the 
image correspond to the foreground objects instead of 
conventional geometry model, BRDF and illuminant. It’s 
obvious that the Environment Matting can process specular 
reflection and refraction excellently and work little on diffusion. 
Therefore, because in close range, the main problem lies in 
photos is the different illuminations make the different 
representations of similar surface of the scene imaging in 
different time, the method can process applications in close 
range photogrammetry to dodge between different illumination 
environments although is should be improved further to take 
refraction into account. 
4. DODGING BASED ON IMAGE INFORMATION 
In the applications of photogrammetry and remote sensing, 
when the scale is large enough to ignore the specular reflection 
of the surface in the scene or the surface can be looked as 
Lambertian, such as the ground in remote sensing and some 
aero-photogrammetry applications, that is to say, when the 
process only takes refraction, diffusion, and absorption into 
account, the task of dodging can be performed only based on 
image information, without calculating the geometry model and 
reflectance. So far, lots of methods in such direction on the 
topic of conventional dodging are proposed. Most of them are 
arranged into 6 typical classes as following. 
(1) Simple Template/Model. In some simple applications, an 
intuitive idea is to build a template to indicate the uneven 
lightness in the image and to modify each pixel’s value based 
on such template [Zhang, 2003a]. Another typical approach is 
to build a simple geometry model based on the characteristic of 
optical lens and to modify each pixel’s value based on the 
geometry model [Milan, 2002]. The advantages of such 
methods are computation cost few, the main defects are some 
parameters should be given before processing and geometry 
quality can't be improved. Additional, the last approach can 
process only one type of lightness distribution. 
(2) Image Restoration. Such approaches consider the solution in 
the view of image restoration completely. They construct 
degradation model based on the cognition that when the 
lightness of the image is balanced, the entropy of the image 
after processing should be maximum [Tian, 1999; Wang, 2000]. 
  
The effect is right, while the defect is computation cost is large. 
In the literature [Tian, 1999], the algorithm is practice in a chip. 
(3) Mask Simulation. Based on the Mask technique in 
traditional photograph, construct a transparent Mask to restore 
the geometry attributions and radiometry attributions 
simultaneously [Hu, 2004]. The effect of balancing lightness 
and color and geometry improvement is excellent. However, its 
dynamic range is not improved enough, some detail is lost, and 
it can’t eliminate the color cast in the image. 
(4) Homomorphic Filtering. Such applications believe that the 
illumination is consists of relative high frequency incident light 
and relative low frequency environment reflect light. 
Homomorphic filtering can balance the lightness through 
restraining high frequency part and improving low frequency 
part [Peli, 1984]. Main defects of such applications arc the lost 
of some detail in low frequency part, compression of the 
dynamic range, and reduced holistic lightness. 
(5) Computational Color Constancy. The improving retinex 
theory is a very important and promising approach [Land, 
1977]. Multi-scale retinex for color rendition (MSRCR) 
[Rahman, 1996] is an effective version of the theory, which can 
provide dynamic range compression, color constancy and color 
rendition simultaneously. Lots of algorithms provide some 
functions of such aspects. But they can't work in all aspects. In 
the depiction of relative literatures [Barnard, 1999], it’s 
convictive that such approach can process large mass of data in 
dynamic range compression, contrast improving, and color 
rendition effectively and it’s a promising approach. However, 
as to traditional dodging, the effect is not as good as the former 
approaches. 
(6) Other Approaches. Besides the approaches above, literatures 
indicate that some other approaches can improve dodging effect 
in some aspects. However, such approaches arc almost the 
development of some algorithm above and the computation is 
complex. Such as the approach based on wavelet, which is a 
development of homomorphic filtering. [Zhang, 2001] The 
former improves the latter aimed at its defect of loss some 
detail in low frequency area. It splits the image into different 
frequency parts and improves the quality of low frequency part 
to keep fine details. 
It’s noticeable that the first kind of method can be looked as a 
simplified version of the Mask Simulation. While the 
Homomorphic Filtering and Mask Simulation can be united into 
a whole frame like the MSRCR because the MSRCR can 
process the image in multi scales and in each scale, it control 
the strength with a Gauss function, which facilitate the unite of 
the form and the improvement. What's more, except the 
MSRCR, other approaches all process color images in three 
channels separately. To obtain a straightforward cognition, à 
grey image is tested by the five approaches above. The test 
results are listed following (Fig.1-Fig.6). 
To display the difference, the parameters of each algorithm in 
processing are not tuned to the best value. The examples here 
are merely for emphasis the main difference on dynamic range 
compression, lightness balancing, and geometry improvement 
in each algorithm. Because the images are zoomed in, the 
description of the compare is listed in Tab.l. Where, L.B. 
means the effect of lightness balancing. 1.R. means the intensity 
range. G.I. means the geometry improving. C.C. means 
computation cost. C.I. means the degree of contrast 
improvement. The number of asterisk means the strength of the 
indicator. 
Inte 
  
Obvi 
proc 
are t 
obtai 
Fig. 
  
Fig.7