Bernal S D_2010.pdf - University of Plymouth

5.1. FEEDFORWARD PROCESSING
5.1.2 Object categorization
This section describes the performance of the model during feedforward categorization, based
on the feedforward processing conslraints defined in Section 4.4. The network waslrained using
60 object silhouette images, shown in I'igure 5.^, from which the S2 and S3 prototypes were
learned. The trained network was then tested on different transformations of the same images
including occluded, translated and scaled versions.
Throughout this section, I have used correct categorization to mean that the state with the
maximum value of the model's lop layer response corresponds to the input image, for the
3-level and 4-leve! architectun;s, the distributions of the four top layer nodes, corresponding
to each of the four S2 RF sizes, are averaged, resulting in a single distribution with 60 states.
Additionally, for the alternative 3-Ievel architecture, the values of the four prototypes learned
for each object category are also averaged, leading again to a single dislribulion with 60 slates.
The model's performance is measured as a percentage of correctly categorized images for each
datasel of 60 images.
For the occluded lest set an average of 30% of the image's black pixels are deleted using a
reclangular while patch. The rectangle is placed in a position thai leaves the image identifiable
to a human observer In the translated test-set. the object is moved to a new position within
the available image frame of 160x160 pixels. The displacement will be near to the maximum
permitted in both directions but will depend on the original object size, i.e. small objects allow
for bigger displacements. Two different scale sets have been used; scale ±10%, where the
image is scaled to either 90% or 110% of the original size and centred; and scale ±20%. where
ihe image is scaled to either 80% or 120% of the original size and centred. An example of the
different transformations for five arbitrary images is shown in Figure 5.9.
200