Segmented face detection using clustering

Show full item record

Redirect: RIT Scholars content from RIT Digital Media Library has moved from to RIT Scholar Works, please update your feeds & links!
Title: Segmented face detection using clustering
Author: Srimanth, Gunturi
Abstract: Perception forms a very important part of learning. The way we perceive things has a lot to do with how we understand. It forms a very crucial link in our build-up of knowledge. Living organisms have a remarkable ability of understanding spatial information. It is due to their inherent ability of generating native organizations, models, etc, and most importantly, their ability to generalize and infer - based upon symmetry, probability, familiarity, etc, that allows them to instantly adapt their knowledge to the given surroundings. It forms a very basic step in survival. When trying to make machines intelligent, one of the first hurdles faced is the problem of perception - questions like which data is important (light, color, texture, sound, etc.)? how much importance should each data be given? etc come up. The purpose of this work is to observe the workings, and results, of trying to detect faces in images, by searching separately for the eyes, nose and mouth regions of a face. The regions are searched independently of one another, using clustering and Neural Networks. Broadly speaking Clustering is used to locate generalized face regions, and Neural Networks are used to map accurately the decision surface. The search for eyes, nose and mouth is done separately, in an attempt to reduce the complexity of the intensity map being searched, thus hoping to improve upon the accuracy and reliability of the detection process. Also it provides for simultaneous parallel search, which is of high importance for real-time tasks like face-detection. To observe the effectiveness, and generalization capability of the process, a very small and localized dataset was used for training pur poses. Also, no feature sets or such abstractions were used in the training and implementation of the work - only raw data was used for detection - this was done to reduce the effects of selection of a wrong feature set.
Record URI:
Date: 2001

Files in this item

Files Size Format View
GSrimanthThesis02-2001.pdf 8.826Mb PDF View/Open

The following license files are associated with this item:

This item appears in the following Collection(s)

Show full item record

Search RIT DML

Advanced Search