Ordinal Shape Coding and Correlation for Orientation-invariant 2D Shape Matching

The human brain and visual system is highly robust and efficient at recognising objects.
Although biologically inspired approaches within the field of Computer Vision are
often considered as state of the art, a complete understanding of how the brain and
visual system works has not yet been unlocked. Benefits of such an understanding are
twofold with respect to Computer Vision: firstly, a more robust object recognition
system could be produced and secondly a computer architecture as efficient as the
brain and visual system would significantly reduce power requirements. Therefore it
is worthy to pursue and evaluate biologically inspired theories of object recognition.

This engineering doctorate thesis provides an implementation and evaluation of a
biologically inspired theory of object recognition called Ordinal Shape Coding and
Correlation (OSCC). The theory is underpinned by relative coding and correlation
within the human brain and visual system. A derivation of the theory is illustrated
with respect to an implementation alongside proposed extensions. As a result, a
hierarchical sequence alignment method is proposed for the correlation of multi-
dimensional ordinal shape descriptors for the context of orientation-invariant 2D shape
descriptor matching.

Orientation-invariant 2D shape descriptor matching evaluations are presented
which cover both synthetic data and the public MNIST handwritten digits dataset.
Synthetic data evaluations show that the proposed OSCC method can be used as a
discriminative orientation-invariant 2D shape descriptor. Furthermore, it is shown that
the close competitor Shape Context (SC) method outperforms the OSCC method when
applied to the MNIST handwritten digits dataset. However, it is shown that OSCC
outperforms the SC method when appearance and bending energy costs are removed
from the SC method to compare pure shape descriptors. Future work proposes that
bending energy and appearance costs are integrated into the OSCC pipeline for further
OCR evaluations.