Patch-based semantic labelling of images.
The work presented in this thesis is focused at associating a semantics to the content of an image, linking the content to high level semantic categories. The process can take place at two levels: either at image level, towards image categorisation, or at pixel level, in se- mantic segmentation or semantic labelling. To this end, an analysis framework is proposed, and the different steps of part (or patch) extraction, description and probabilistic modelling are detailed. Parts of different nature are used, and one of the contributions is a method to complement information associated to them. Context for parts has to be considered at different scales. Short range pixel dependences are accounted by associating pixels to larger patches. A Conditional Random Field, that is, a probabilistic discriminative graphical model, is used to model medium range dependences between neighbouring patches. Another contribution is an efficient method to consider rich neighbourhoods without having loops in the inference graph. To this end, weak neighbours are introduced, that is, neighbours whose label probability distribution is pre-estimated rather than mutable during the inference. Longer range dependences, that tend to make the inference problem intractable, are addressed as well. A novel descriptor based on local histograms of visual words has been proposed, meant to both complement the feature descriptor of the patches and augment the context awareness in the patch labelling process. Finally, an alternative approach to consider multiple scales in a hierarchical framework based on image pyramids is proposed. An image pyramid is a compositional representation of the image based on hierarchical clustering. All the presented contributions are extensively detailed throughout the thesis, and experimental results performed on publicly available datasets are reported to assess their validity. A critical comparison with the state of the art in this research area is also presented, and the advantage in adopting the proposed improvements are clearly highlighted.
- Theses