Abstract:
This thesis describes a new technique for Spacio-Spectral Object Detection. The image
is translated into its spacio-spectral information, which is 3 dimensional in case of grey
scale image while for colored image its dimension depends upon the number of color
channel used. The idea is to model the environment (background) as jointly Gaussian
random variables, which leads to fitting an ellipsoid in case of grey scale image.
Colored images with 2 color channels and 3 color channels are also modeled using the
same technique, which leads to fitting of closed surfaces in higher dimensions. The
highly dense clustered data present within the boundaries of ellipsoid or closed surfaces
(in higher dimensions) are termed as environment and those data points outside the
boundaries are named as outliers. The dense data cluster laying in outliers form an
object and the number of objects present is also determined. Simulations are performed
on images with different sizes of objects and different kind of color contrasts, which
reveal the effectiveness of this algorithm. A comparison is made between
spectrum(single dimensional data which only uses intensity of image) and spaciospectral object detection, which shows that significantly better results are achieved
using the newly developed technique. The results also manifest that as the dimension
increases; the detection becomes more and more accurate. Thus this novel technique
successfully detects objects both in grey scale and colored images.
