Automatic Location of Human Faces by Machine

Robertson, Graham (1994) Automatic Location of Human Faces by Machine. PhD thesis, University of Glasgow.

Full text available as:
[thumbnail of 13833805.pdf] PDF
Download (26MB)


This thesis investigates and develops new methods for locating a human face in a computer image. The need for such a technique has long been recognised in many automatic face processing applications such as face recognition or facial image compression. Other studies have attempted face location and the related task of face feature location, with varying degrees of success, on face images captured in controlled conditions; so that the face is on a plain background or is of a fixed size. The result of this study has been the development of a location system that can cope with background clutter, variable size images, slightly rotated faces and changes in overall lighting. This improvement in face location technique has been achieved without an increase in processing time. The overall system concurrently finds possible eyes, noses and mouth in the image and a control system combines these feature proposers into the complete face location system. The first part of the system consists of robust preprocessors based on previously identified intensity signatures that indicate the possible presence of the desired facial features. Roughly, these signatures are continuous vertical troughs in intensity for noses, continuous horizontal troughs in intensity for mouths and multi-direction troughs in intensity for the eyes. These preprocessors reduce the image search space without rejecting the true facial features being searched for. This reliable reduction in search space is the main contributor to the efficiency of the entire system. The second part of the system analyses the intensity gradient directions on features using statistical methods that were previously trained on many manually located facial features. This technique, called the PRODIGY, combines probability density functions of real and false features in a likelihood function. This function gives an output for each feature pertaining to the likelihood of it being a real facial feature or a distracter (false) feature. The control system combines the possible features with a statistically controlled spatial model of how the individual features are related to each other. It then selects the most likely set of proposed features as the location of the face in the image. The overall system proved to be successful on a high proportion of face images and the usefulness of the technique is demonstrated by a simple but effective face recognition system.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Additional Information: Adviser: Brian Rosner
Keywords: Electrical engineering, Computer engineering, Artificial intelligence
Date of Award: 1994
Depositing User: Enlighten Team
Unique ID: glathesis:1994-75678
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 19 Nov 2019 18:58
Last Modified: 19 Nov 2019 18:58

Actions (login required)

View Item View Item


Downloads per month over past year