Date of Award


Degree Type


Degree Name

Doctor of Philosophy


Engineering and Mathematics


Faculty of Computing, Health and Science

First Advisor

Associate Professor Abdesselam Bouzerdoum

Second Advisor

Dr Douglas Chai


Automatic human face detection in digital image has been an active area of research over the past decade. Among its numerous applications, face detection plays a key role in face recognition system for biometric personal identification, face tracking for intelligent human computer interface (HCI), and face segmentation for object-based video coding. Despite significant progress in the field in recent years, detecting human faces in unconstrained and complex images remains a challenging problem in computer vision. An automatic system that possesses a similar capability as the human vision system in detecting faces is still a far-reaching goal. This thesis focuses on the problem of detecting human laces in color images. Although many early face detection algorithms were designed to work on gray-scale Images, strong evidence exists to suggest face detection can be done more efficiently by taking into account color characteristics of the human face. In this thesis, we present a complete and systematic face detection algorithm that combines the strengths of both analytic and holistic approaches to face detection. The algorithm is developed to detect quasi-frontal faces in complex color Images. This face class, which represents typical detection scenarios in most practical applications of face detection, covers a wide range of face poses Including all in-plane rotations and some out-of-plane rotations. The algorithm is organized into a number of cascading stages including skin region segmentation, face candidate selection, and face verification. In each of these stages, various visual cues are utilized to narrow the search space for faces. In this thesis, we present a comprehensive analysis of skin detection using color pixel classification, and the effects of factors such as the color space, color classification algorithm on segmentation performance. We also propose a novel and efficient face candidate selection technique that is based on color-based eye region detection and a geometric face model. This candidate selection technique eliminates the computation-intensive step of window scanning often employed In holistic face detection, and simplifies the task of detecting rotated faces. Besides various heuristic techniques for face candidate verification, we developface/nonface classifiers based on the naive Bayesian model, and investigate three feature extraction schemes, namely intensity, projection on face subspace and edge-based. Techniques for improving face/nonface classification are also proposed, including bootstrapping, classifier combination and using contextual information. On a test set of face and nonface patterns, the combination of three Bayesian classifiers has a correct detection rate of 98.6% at a false positive rate of 10%. Extensive testing results have shown that the proposed face detector achieves good performance in terms of both detection rate and alignment between the detected faces and the true faces. On a test set of 200 images containing 231 faces taken from the ECU face detection database, the proposed face detector has a correct detection rate of 90.04% and makes 10 false detections. We have found that the proposed face detector is more robust In detecting in-plane rotated laces, compared to existing face detectors. +D24