Résumé:
The objective of this work is to design and develop a 2D shape descriptor for the
recognition, matching and shape retrieval, more robust and efficient than other existing
descriptors. The proposed approach introduces a new primitive of a geometric nature
called the shape context based on the feature points extracted from the contour, which
makes it possible to represent and characterize the shape in a simple and more reduced
way.
The shape is given by a contour representation which combines the critical points and
the sampled points extracted from the contour to form a vector of feature points. While
each point of characteristic is associated with a log-polar histogram of directions and
distances relative to this point. The set of histograms associated with all the characteristic
points forms what is called the SC descriptor.
To overcome the shape matching problem, a matching algorithm is designed to compare
a query shape with a collection of target shapes. The result of the matching between each
pair of shapes is stored in a cost matrix.
We validated the proposed approach using a base of 2D shapes. The test results
obtained are favorable and promising and show that the new approach is competitive with
other approaches for shape matching and retrieval tasks and has a descriptor that is more
robust to noise, occlusions and invariant to geometric transformations.
