Abstract:
Edges are important features of an image which contain only signi cant information con-
tent and discard remaining irrelevant features such as background. Over the years, re-
searchers have been thriving to come up with new methods and techniques in edge de-
tection. Classical edge detectors target the abrupt change in intensity values of pixels in
a given image. The location of this change in intensity is marked as an edge. The major
problem in those methods was lack of continuity in edge lines, detection of false edges,
and missing out the corners. This thesis proposes an edge detection algorithm which is
based on Smallest Univalue Segment Assimilating Nucleus (SUSAN) area principle, which
is comparatively a new approach by creating a circular mask (square matrix with zero
padding at the corners) to scan an image pixel-by-pixel. SUSAN area for each pixel is
calculated and an image is transformed into its corresponding SUSAN area matrix. New-
ton's law of gravity is then applied to the SUSAN area matrix by treating the values of
elements as mass of an object. The gravitational force exerted by an element on its neigh-
bors is calculated, and non-zero gravity magnitude is then compared with an appropriate
threshold to declare an element as an edge or no edge. A variety of battle eld images are
selected to test the performance of the proposed algorithm. Results obtained from this
algorithm are compared with modern edge detecting algorithms, so that an assessment
can be carried out. To assess, the results are segmented into sub parts and human judg-
ment and estimate has been performed over all these segments. The estimate is further
quanti ed by evaluating the percentage of correct edge detection performed by all the
algorithms. Further, quantitative analysis are performed to bring out the detection capa-
bilities of the proposed algorithm with others. It is observed from the results presented
in chapter 4 that the proposed algorithm has performed better than other algorithms in
terms of edge integrity, corner detection, no loss of information and no false detection.
