Scene Segmentation Using Polarisation Information

This thesis draws its hypothesis from the quantitative description of reflection and refraction
of light at interfaces, given by the Fresnel theory. According to the theory, initially
unpolarised light is partially polarised upon reflection from a smooth dielectric surface.
Hence, variations in the polarisation properties of reflected and scattered light must indicate
a change in properties of the reflecting surface. Assuming internal changes in refractive
index can be neglected, any substantial change in the polarisation of light is thus
indicative of a change in the material remitting the light. The contribution of this thesis
is to develop a method for image segmentation based on surface material characteristics.
The novel aspects of the method are the expression of intensity distribution as a function of
the surface zenith and azimuth angles, and the expansion of this function using spherical
harmonics to estimate surface characteristics. The method begins with estimating shape
from polarisation using the work of Atkinson and Hancock. The surface normals obtained are used in combination with
pixel intensities to form a three dimensional function that describes intensity changes in
the image. This function is then expanded as a harmonic series and the constants of the expansion
are used as features to characterize image regions and to segment the image based
on the hypothesis. Experimental evidence is presented through analysis of polarisation
images and segmentation of images of different objects.