Harris, Christopher; (1997) An Investigation into the Application of Genetic Programming Techniques to Signal Analysis and Feature Detection. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text An_investigation_into_the_appl.pdf Download (9MB) | Preview

Abstract

This thesis investigates the use of the Genetic Programming (GP) machine learning paradigm to computer vision problem domains related to the detection of features within discrete data signals. The use of GP to find a solution to such problems involves two pieces of design. The first is the set of elements that the search process uses to construct possible problem solutions, known as the primitive set. The second is the design of an appropriate reward mechanism for candidate solutions, known as the fitness function. The wide range of signal types and features to be detected in computer vision detection problems ensures that representation of both data and solutions is of prime importance. Two contributions made by this thesis are to provide approaches to both primitive set design and fitness function design that are generic to feature detection problems. A three-component fitness function design is presented which reflects the desirable properties of a generic feature detector. Work is also presented on a design for a primitive set that is applicable to a wide range of signal processing problems. These techniques are explored using the classic problems of edge detection and template matching as experimental test-beds. Work in this thesis on edge detection produces filter functions that outperform those produced by human experts under real-world conditions. This process can be used to produce edge detectors optimised for specific sets of data. Working from the basis that Strongly Typed Genetic Programming (STGP) is essential for solving vision problems, a fourth contribution is the adoption of STGP as a general syntactic constraint mechanism for the production of GP program trees. This provides a structuring mechanism in addition to allowing the use of complex and relevant data types within candidate programs. By using the structuring to enforce a hierarchy of abstraction in terms of data types and representations, we can increase the power of GP to solve hard problems, and allow more intelligent use of the limited search power available with finite computational resources. This is a form of abstraction not previously used in GP.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item