Use this URL to cite or link to this record in EThOS:
Title: Superpixel lattices
Author: Moore, A. P.
ISNI:       0000 0004 2731 9134
Awarding Body: University College London (University of London)
Current Institution: University College London (University of London)
Date of Award: 2011
Availability of Full Text:
Access from EThOS:
Access from Institution:
Superpixels are small image segments that are used in popular approaches to object detection and recognition problems. The superpixel approach is motivated by the observation that pixels within small image segments can usually be attributed the same label. This allows a superpixel representation to produce discriminative features based on data dependent regions of support. The reduced set of image primitives produced by superpixels can also be exploited to improve the efficiency of subsequent processing steps. However, it is common for the superpixel representation to have a different graph structure from the original pixel representation of the image. The first part of the thesis argues that a number of desirable properties of the pixel representation should be maintained by superpixels and that this is not possible with existing methods. We propose a new representation, the superpixel lattice, and demonstrate its advantages. The second part of the thesis investigates incorporating a priori information into superpixel segmentations. We learn a probabilistic model that describes the spatial density of object boundaries in the image. We demonstrate our approach using road scene data and show that our algorithm successfully exploits the spatial distribution of object boundaries to improve the superpixel segmentation. The third part of the thesis presents a globally optimal solution to our superpixel lattice problem in either the horizontal or vertical direction. The solution makes use of a Markov Random Field formulation where the label field is guaranteed to be a set of ordered layers. We introduce an iterative algorithm that uses this framework to learn colour distributions across an image in an unsupervised manner. We conclude that our approach achieves comparable or better performance than competing methods and that it confers several additional advantages.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available