Wright, Adrian; (2001) Image Segmentation using Local Surface Fitting. Doctoral thesis (Ph.D), UCL (University College London).

Text Image segmentation using local surface fitting.pdf Download (13MB)

Abstract

Images contain information and the aim of digital image processing is generally to make the extraction of this information easier or even to automate it. Segmentation is the division of digital images into regions. The final aim is usually scene segmentation in which the regions correspond to actual objects in the image, for example a car, a brain tumour, a flooded area. A more fundamental process is image segmentation, the subject of this thesis, which is entirely context free and results in regions which are homogeneous in themselves but may not necessarily correspond to whole objects. The basic assumption here is that all images can be segmented into regions that have certain consistent characteristics. A region classification is defined which postulates two basic region features, a 'smooth' grey level variation overlayed by textural variations which include 'noise'. This thesis is concerned with the 'smooth' grey level variations and the problems they present for segmentation. A novel method of calculating the degree of connectivity of neighbouring pixels is presented based on the similarity of their best fitting local surfaces. The strength of a segmentation is defined as a sum, taken over all pixel-neighbour pairs, of these connectivities. The 'best' segmentation is defined as that with the greatest strength and it is shown how this definition parallels the intuitive idea of a good segmentation by completing boundaries even where they are locally weak. Methods of approaching this defined 'best' segmentation are described starting with a simple thresholding of the connectivities and proceeding to optimisation using simulated annealing. Results using simulated and real images are presented and informally compared with a widely used segmentation technique based on Markov Random Fields.

Download activity - last month

Export as CSVJSONXML

Download activity - last 12 months

Export as CSVJSONXML

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item