BibTeX entry

@PHDTHESIS{200312Niklas_Ludtke,
  AUTHOR={Niklas Ludtke},
  TITLE={A Population Coding Approach to Edge Detection
    and Perceptual Grouping},
  SCHOOL={University of York},
  MONTH=Dec,
  YEAR=2003,
  URL={http://www.bmva.org/theses/2003/2003-ludtke.pdf},
}

Abstract

This dissertation presents a novel, biologically inspired approach to edge detection and perceptual organisation, based on a synthesis of the well-known Gabor filters with the concept of population coding from computational neuroscience. A Gabor filter bank is regarded as an ensemble of orientation sensitive units that encode local contour orientation in a distributed fashion, somewhat akin to the “simple cells” in the mammalian primary visual cortex. From the filter ensemble, a probability density function (pdf) of local contour orientation is decoded by taking into account the orientation tuning function of the filters and assuming a von Mises mixture model for the contour angle. The parameters of the pdf are estimated using an expectation maximisation (EM) algorithm. Whereas conventional edge detection schemes tend to reduce the set of filter responses in each pixel to a single quantity, e.g. a local tangent angle, this dissertation takes a different approach, aiming to maintain a distributed representation. The benefits of the resulting analytically derived probabilistic population decoding algorithm is that points with multiple orientations, such as corner points or junctions, can be accommodated within the same framework by means of multimodal probability densities. Another important aspect of distributed coding is the notion of certainty, characterised by the spread of activity across the filter bank or the entropy of the orientation pdf. It is demonstrated that the availability of local feature certainty prior to perceptual organisation is beneficial for feature localisation. Selecting features by means of their certainty, rather than by thresholding filter responses, renders the feature extraction contrast independent and more robust against noise. In the subsequent grouping step, small curved contour segments are generated through spline interpolation between pairs of locally extracted tangent elements. The grouping process involves a revision of the local orientation measurements, controlled by their certainty values and the overall curvature of the connecting spline. This possible only because certainty has become a measured quantity determined at the stage of local extraction. In most other grouping schemes, certainty is either not considered, or, as in probabilistic relaxation labelling, inferred after feature extraction during the optimisation of probability density parameters representing local features. While not claiming to present a model of biological visual processing, this thesis provides some new insight into the initial problems that both artificial and neural visual systems are confronted with: the extraction and representation of local features following sensory acquisition, and the subsequent grouping of such locally extracted features into larger, more complex entities.