PhD Thesis

 

Defended on 26th April, 2013

COSFIRE (Combination of Shifted Filter Responses): A Trainable Filter Approach to Visual Pattern Recognition

Propositions of the PhD thesis

  1. The CORF (Combination of Receptive Fields) model of a simple cell outperforms the Gabor function model
  2. A nonlinear, multiplicative response of a CORF model simple cell is biologically more plausible than linear spatial summation
  3. Simulated reverse correlation is a powerful tool to determine the properties of a filter
  4. An adaptive and irregular area of support of a COSFIRE filter allows the modelling of complex visual features and shapes
  5. The shape of an object can be effectively represented in a hierarchical way by the mutual spatial arrangement of simple lines and edges, followed by vertices, combinations of vertices, etc …
  6. A blurring function based on a weighted maximum rather than weighted sum is more robust to allow for some tolerance in the relative locations of contour parts of a pattern of interest
  7. A COSFIRE filter is selective to the shape properties of a pattern of interest and it is more robust to contrast variations and presence of texture than methods that use (dis)similarity measures to compare descriptors
  8. Effectiveness has priority over efficiency
  9. The rejection of a paper should not be considered as a setback, but an opportunity to improve the work and to re-submit it to a more reputable journal

Johann Bernoulli Institute for Mathematics and Computer Science, University of Groningen, the Netherlands