|Facial image analysis for video indexing and retrieval
|Hong Kong Polytechnic University -- Dissertations.
Human face recognition (Computer science)
Image processing -- Digital techniques.
|Department of Electronic and Information Engineering
|xiv, 126 p. : ill. (some col.) ; 30 cm.
|The aim of this research is to investigate efficient schemes for facial image analysis in video retrieval and indexing. Statistics have shown that over 95% of the primary camera's subjects in videos are humans, therefore face analysis in videos can greatly benefit on video retrieval and indexing. Our research focuses on three areas: face detection, face recognition, and indexing. Some popular techniques and recent developments of the methods for both face detection and recognition are also reviewed. In this project, we have proposed an effective template, namely Spatially Maximum Occurrence Template (SMOT), for face detection. This template is combined with a mixture of Gaussian models to verify whether an image region is a face or not. SMOT has a high representative power for faces, and can detect faces under various conditions. We have also proposed an efficient method for face recognition. A simplified version of the Gabor wavelets (SGWs) has been devised for feature extraction. Gabor wavelets (GWs) have commonly been used for extracting local features which are insensitive to environmental factors, but extracting these features is computationally intensive. Simplified Gabor wavelets (SGWs) are therefore devised, and an efficient algorithm for extracting the features based on an integral image is proposed. These SGW features are then applied to face recognition. Experiments show that using SGWs can achieve a performance level similar to that using GWs, and the runtime for feature extraction using SGWs is 4.39 times faster than that of GWs implemented by using the fast Fourier transform. An efficient indexing structure for searching face images in a large database has also been investigated and proposed. This indexing structure is formed by a number of vantage objects, which are constructed using the discriminative features extracted from Gabor wavelets. The training faces in a large database are ranked in order with reference to each of the vantage objects, so a ranked list is constructed for each vantage object. A query face image will also be ranked with respect to each vantage object, and those neighboring training faces to the query face in the respective ranked lists are selected to form a much smaller database, called a condensed database. Experiments show that a condensed database whose size is 25% of the original large database can be formed with a probability of 99.3% that the matched face to the query input exists in the condensed database. Then, a more computational and accurate recognition algorithm can be adopted in the condensed database without any degradation of the recognition accuracy.
|All rights reserved
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item: