Connectionist Schemes for Image Segmentation.

Date of Submission

December 2000

Date of Award

Winter 12-12-2001

Institute Name (Publisher)

Indian Statistical Institute

Document Type

Master's Dissertation

Degree Name

Master of Technology

Subject Name

Computer Science

Department

Electronics and Communication Sciences Unit (ECSU-Kolkata)

Supervisor

Pal, Nikhil Ranjan (ECSU-Kolkata; ISI)

Abstract (Summary of the Work)

The technique of extracting information from an image is referred as inage analysis. Generally, the first step of image analysis is to negment the image into ita constituent parta or objecta. Autonomous segmentation is one of the most difficult tasks in image proceesing. Fu and Mui [7] categorined seginentation techniques into three classes. (1) feature based elustering, (2) edge detection and (3) region extraction. Segmentation is a procees of partitioning the image into BOme non intersecting regions such that each region is nearly homogeneous. Formally it can be defined (8] as follows : if F is the set of all pixels and p() is the uniformity (homogeneity) predicate defined on groups of connected pixels, then the segmetation is a partitioning of the set of connected subeets or regions (S1...Sn) such that uin=1si=Fsiusj=o, i=jwithSina, = 4, ij The uniformity predicate p(si) - true for all regions a, and p(si u si) is false, when a, i adjacent to ay. A large number of seginentation techniques are present in the literature, bu there is no single method which can be considered good for all images, also all methods ar not equally good for a particular type of linage. There are many challenglng ina ues like, the developinent of a unified approach to image segmentation which can (probably) be applied to al kinds of imagen. Till now, there in no universally accepted method of quantiication of segmentec output. Authentication of edges is alao a very important task. Different edge operators like Sobel, Prewitt, Mar-Hiklreth, ete. compute a munerical value using its edge operator(s) at every pixel location to indicate whether an edge is present or not at that location. Here this numerica value may be termed as a measure of edginesn at that pixel location. However, all of them are not valid candidates for edges. Normally a threshold needa to be applied on the edginess to ge the edge pixela. The selection of a suitable threshold in very important and a dificult tank. I is crucial because for some part of the image, low intenaity variation may correspond to edge of interest, while the other part may require the high intensity variation. Adaptive thresholding (910] (11) often in taken as A solution of thin. Obviounly It cannot elininate the problem o threslold selection. A good atrategy to produce ineaningful segnenta would be to fuse region segmentation results and edge outputa.One may attempt to extract the segments in a variety of ways. Broadly, there are twc approaches namely, classical approach and fuzzy mathematical approach. Under the claasica. approach we have seginentation techniques based on histogram thresholding, edlge detection relaxation, and semantic and syntactic approaches. In aklition to these there are some methode which do not fall clearly in any one of the above clannes. Sinilarly, the furzy mathemnatical approach also has methoda baned on edge detection, threeholding, and relaxation. Some of these methods, particularly the histogram based methods are not all suitable for noiny imnnges. Several attempta have alno been made to develop image procersing algorithin uning Artificial Neural Network (ANN) models, particularly Hopfiled and Kohonen Neural Network models. These algorithms work well even in a highly noisy environment and they are capable of producing outputs in real time applications. Since the proposed methods are neural net based we provide a brief summary of some NN based segnentation techniques.

Comments

ProQuest Collection ID: http://gateway.proquest.com/openurl?url_ver=Z39.88-2004&rft_val_fmt=info:ofi/fmt:kev:mtx:dissertation&res_dat=xri:pqm&rft_dat=xri:pqdiss:28843062

Control Number

ISI-DISS-2000-141

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

DOI

http://dspace.isical.ac.in:8080/jspui/handle/10263/6311

This document is currently not available here.

Share

COinS