The British Machine Vision Association and Society for Pattern Recognition 

BibTeX entry

  AUTHOR={Ian Scott},
  TITLE={Searching Image Databases Using Appearance Models},
  SCHOOL={University of Manchester},


This thesis examines the problem of retrieving images from large image databases by detecting objects of interest. The approach adopted involves combining model-based recognition, using Active Appearance Models (AAMs), with a sophisticated statistical classifier. An experimental evaluation of two published methods for achieving very high detection accuracy leads to the choice of a Support Vector Machine (SVM) classifier with iterated negative-example refinement. The method for training the SVM is further developed, leading to a fully-automated approach to choosing training parameters. Similarly, the standard AAM approach is extended, leading to significant improvements in performance. One contribution is the “Texture AAM,” which replaces the grey-level values, ordinarily used in the AAM’s shape-normalised patch, with non-linear descriptors of local edge and corner structure. Another contribution is to investigate, more fully than previously, the effects of limiting the AAM parameters during image search – leading to a reappraisal of the optimal approach. Finally, the improved AAM and SVM are combined to create a novel “AAM-SVM” system for image retrieval, that is shown to be significantly more effective than either method alone. Extensive experiments are performed to analyse the behaviour of the system, demonstrating that detection accuracy is superior to that of a simple patch-based SVM approach that is among the state-of-the-art methods. The approach is, however, many times too slow for practical applications. This leads to an initial investigation of a multi-stage approach, which uses an AdaBoost patch classifier as a first stage. The speed of this system approaches that necessary for practical image database search.