Until now texture has been largely viewed as a statistical or holistic paradigm: textures are described as a whole and by summary statistics. In this thesis it is assumed that there is a structure underlying the texture leading to models, reconstruction and to scale based analysis. Local Binary Patterns are used throughout as the basis functions for texture and methods have been developed to reconstruct texture images from arrays of their LBP codes. The reconstructed images contain identical texture properties to the original; providing the same array of LBP codes. An evidence gathering approach has been developed to provide a model for each texture class based on the spatial structure of these elements throughout the image. This method, called Evidence Gathering Texture Segmentation, provides good results for segmentation with smooth boundaries and minimal oversegmentation, when compared with existing methods. Analysing microand macro-structures confers ability to include scale in texture analysis. A novel combination of lowpass and highpass filters produces images devoid of structures at certain scales; allowing both the micro- and macro-structures to be analysed without occlusion by other scales of texture within the image. A two stage training process is used to learn the optimum filter sizes and to produce model histograms for each known texture class. The process, called Accumulative Filtering, gives superior results compared to the best multiresolution LBP configuration and analysis only using lowpass filters. By reconstruction, by evidence gathering and by analysis of micro- and macro-structures, new capabilities are described to exploit structure within the analysis of texture.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:605763 |
| Date | January 2014 |
| Creators | Waller, Ben |
| Contributors | Nixon, Mark |
| Publisher | University of Southampton |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | https://eprints.soton.ac.uk/364755/ |
Page generated in 0.0023 seconds