UCL Discovery
UCL home » Library Services » Electronic resources » UCL Discovery

Contextual image classification

Poole, Ian; (1992) Contextual image classification. Doctoral thesis (Ph.D), UCL (University College London). Green open access

[img] Text
Contextual_image_classificatio.pdf

Download (11MB)

Abstract

This thesis studies some of the practical and theoretical issues arising in the supervised contextual classification of image pixels. The main practical contribution is a learning/classification system, named 'Lapwing'. The system aims to: learn to exploit spatial characteristics such as texture, edge and line; learn to exploit contextual dependencies between classes; be efficient at classification time; return 'honest' probabilistic assessments of classification confidence at each pixel; exploit a parallel SIMD processor for classification. The method involves the use of a genetic algorithm to search for 'good' partitions in a high dimensioned pattern space, the partitions being built up hierarchically as a probability tree. This tree may then be used to produce a class probability image from similar test data. Probabilistic relaxation labelling (PRL) is a popular method of iteratively refining probabilistic assessments in the light of contextual constraints. A study of the theoretical limits of excellence for any PRL scheme is presented. The main result is that no scheme can produce assessments equivalent to those conditioned on all the data in the image. It is also shown that for such a scheme to be optimal the updating function must be tailored for each iteration and that after the first iteration the function will depend on the actual distributions of the raw data, not simply on the spatial correlation of the classes. A new form of PRL - trained probabilistic relaxation (TPR) is presented that uses Lapwing to estimate directly the updating function for each iteration in a particular domain of application. Lapwing is demonstrated on noisy synthetic texture discrimination, edge detection and on problems encountered in remote sensing and medical imaging, with encouraging results.

Type: Thesis (Doctoral)
Qualification: Ph.D
Title: Contextual image classification
Open access status: An open access version is available from UCL Discovery
Language: English
Additional information: Thesis digitised by ProQuest.
Keywords: Applied sciences; Image pixels
URI: https://discovery.ucl.ac.uk/id/eprint/10107624
Downloads since deposit
11Downloads
Download activity - last month
Download activity - last 12 months
Downloads by country - last 12 months

Archive Staff Only

View Item View Item