 Background
It is well known that many ‘image textures’ are functions of illumination
conditions and surface relief. Such changes in appearance
can cause catastrophic failures in image-based texture classifiers.
For instance rotation of the physical
texture sample under fixed illumination conditions can cause significant
changes to its appearance - causing catastrophic failure of classifiers
designed to cope with image rotation. An example is shown below.
Two images of the same directional 3D rotated surface texture with
identical illuminant. The surface has been rotated through of 0 and
90 degrees (indicated by the white arrows in the centre). The illuminant
tilt is kept constant at 0 degrees (indicated by the black arrows in
white circles).
These dramatic changes in image texture can also
be controlled and used advantageously. The appearance of three-dimensional
texture
is enhanced by the use of low level side-lighting. Particular
texture directions may be isolated by controlling the direction
of the illumination.
Overhead, or flat illumination may be used to enhance 'painted'
textures and subdue those due to surface relief variation.
Multiple images
taken
under distinct illumination conditions may be used to extract
information on both on surface relief and surface reflectance
characteristics
using photometric stereo techniques.
Despite the potential pitfalls
and advantages described above very
little work has been published on issues concerning three-dimensional
texture and illumination for texture classification purposes.
More specifically, for automated texture analysis:
- there are few published techniques that can compensate for known variations
in illuminant direction,
- there is little information on how the different surface texture
properties (of three-dimensional relief and reflectance function)
could be exploited
for surface classification or defect detection, and
- there are few techniques that deal with rotation
of the texture sample (rather than rotation of the texture image).
|
