The Human Visual Cortex: Universal
Feature Descriptors and Object Recognition with the GPU
Abstract
A
biologically motivated method of extracting image features is presented
and compared to existing techniques, both of biological motivation and
otherwise. Image features are computed using Gabor
filterbanks. A
series of image processing modules are presented with application to
object recognition, motion estimation and depth maps to demonstrate the
universal nature of the features extracted. This type of
processing
covers both main pathways in the Human Visual Cortex. The
functionality of each module directly maps to an associated region of
the Human Visual Cortex and the processing done in each module
corresponds to the processing known to be done in that region of the
brain, as measured by Functional Magnetic Resonsance Imaging (fMRI)
studies. This work also abstracts the modern Graphic Processing
Unit
(GPU) for biological image processing. This is done primarily in
the
form of SIMD image processing using OpenGL and demonstrates a
noteworthy speedup over the SISD equivalent.