DE-SHADING: INTEGRATED APPROACH TO PHOTOMETRIC MODEL,
SURFACE SHAPE AND REFLECTANCE PROPERTIES
Xiuguang Zhou, Egon Dorrer
Inst. f. Photo. u. Karto., Universität der Bundeswehr München, Germany
ISPRS Commission , Working Group WG III/2
KEY WORDS: Cartography, Surface, Understanding, Research, Experiment, Computer Vision, Photometric Model,
Image De-shading System.
ABSTRACT:
A de-shading problem is presented in this paper. By using a brightness image and its associated height image, the de-shading
problem is stated by estimating the approximate photometric model, the surface reflectance properties and an improved precision of
the given height image. Proposed is a de-shading system. It contains a training frame and a working frame. In the training frame, a
probing algorithm is proposed to determine the approximate photometric model amongst some candidate models. In the working
frame, a region growing algorithm based on least square fitting is proposed to determine the inhomogeneous surface reflectance
properties and a shape from shading algorithm is applied to improve the precision of the given height image. Synthetic images
generated by using the Lambertian model and the Torrance-Sparrow model were used as test images in the experiments. The results
are given to illustrate the usefulness of our approach.
1. INTRODUCTION
There are some common interesting topics in computer vision,
remote sensing, photogrammetry, cartography and their relative
communities, such as shaded-relief, surface reflectance
properties, photometric model, the direction of the light source
and shape from shading (SFS). As is well known, shaded-relief
(shading) is to illuminate a surface by using a given light source
or multiple light sources (Brassel, 1973; Horn, 1982; Zhou and
Dorrer, 1995). The surface reflectance properties are important
to study material properties. This is of interest in remote
sensing for observing Earth and the planets. Recently, the
surface reflectance properties have been determined by using
the range and brightness data (Bibro and Snyder, 1988; Ikeuchi
and Sato, 1991; Kay and Caelli, 1994). Shade recovery is a
classic problem in computer vision. One of the techniques to
recover shape is shape-from-shading, which deals with the
recovery of shape from a gradual variation of shading in the
image (Ikeuchi and Horn, 1981; Pentland, 1984; Brooks and
Horn, 1985; Lee and Rosenfeld, 1985; Zheng and Cellappa,
1991; Kimmel and Bruckstein, 1995). There exists quite a
number of photometric models, such as the widely used
Lambertian model, the famous Torrance-Sparrow model
(Torrance and sparrow, 1967) and the Phong model (Phong,
1975). These models are used to describe reflectance maps.
The so-called de-shading in this paper deals with the above
topics. Briefly, de-shading is to remove the natural illumination
from an image to obtain the original information of the object in
the image. It is the inverse procedure of shading. As known, a
shading procedure is to generate an illuminated image by using
the given light source, photometric model, albedo or surface
reflectance properties and the height image (digital terrain
model, DTM). Inversely, if one has an image (maybe a remote
sensing image) and its associated height image (maybe a DTM),
the following questions might be interesting. What is the
approximate photometric model of the image? Where is the
light source for the image? What are the reflectance properties
of the surface? How to increase the precision of the existing
DTM if it is not accurate enough? De-shading tries to solve
these problems.
De-shading is very useful in different application fields such as
computer vision, remote sensing, photogrammetry and
cartography, etc. In the area of remote sensing, e.g., as more
and more DTMs are being successfully generated, one may
want to use the DTM to study the surface properties of the
Earth or other planets. Due to some inadequate conditions (e.g.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
inadequacy of matching algorithm, insufficient information in
the shadow region or errors of the interpolation, etc.), the
precision of the DTM may be insufficient. Therefore, to
increase the precision of an existing DTM is of great important.
Also, one may want to mosaic two remote sensing images with
different directions of illumination. For this problem, we need
first to remove the illuminations of both images (de-shading),
then re-shade the de-shaded images with an assigned
illumination direction based on the obtained proper photometric
model, reflectance properties and improved precision of the
DTM.
2. CONCEPTION AND DEFINITION
As is known, the visual brightness image is the signal recorded
from one or more sensor(s). The sensor receives the visual light
reflected from the surface of the object. The reflected light
comes from the source light which strikes the surface. If the
illumination of an image is removed, what will remain?
Roughly, there will be nothing to be seen. Because no light
source means no visual information. But if we consider the
information recorded on an image, there should be something
"hidden" under the illumination. The information of a visual
image may contain: the direction and energy of the light source,
the reflectance properties of the surface, the geometric
information of the surface, the photometric model information,
the atmospheric affecting information, the noise information
and so on. Obviously, even if the illumination were taken out,
some image information still exists. In other words, some of the
information hidden in the visual grey values is possible to be
estimated. Of course, it is very difficult to get some of the
information listed above (may not be possible to obtain if there
are not enough additional conditions). We named the process of
obtaining some of the hiding image information as de-shading.
In the following, the definition, the task and the inputs-outputs
of the de-shading are given.
Definition of de-shading: Remove the natural expressive
illumination from a visual image to obtain the original
information of the object and the information in the imagery.
Task of de-shading: Given a real image and its associated
approximate height image, the task of de-shading is to obtain
the photometric model approximating the real image, the
albedo or surface reflectance properties, the direction of the
light source and the improved height image which has a higher
precision than the approximate input height image.
1028
Inputs and
some candid
associated a
image). Th:
model appro
for deter
propertie
output the
photometri
photometri
$1 (x, y),
updated he
and gradie
A de-shad
de-shadin;
de-shading
light sour
the appro?
image and
this is not
more or le
fitting me
properties
is used tc
(Ikeuchi ¢
1985; Lee
Kimmel a
strong pr
reflectanc
known ar
over the e
The SFS
propertie
Unfortun
