MULTI-FEATURE BASED FACE DETECTION Qianni Zhang, Ebroul Izquierdo Queen Mary, University of London Mile End Road, E1 4NS, London, UK {qianni.zhang, ebroul.izquierdo}@elec.qmul.ac.uk Abstract In this paper a two-step face detection technique is proposed. The first step uses a conventional skin detection method to extract regions of potential faces from the image database. This skin detection step is based on a Gaussian mixture model in the YCbCr colour space. In the second step faces are detected among the candidate regions by filtering out false positives from the skin colour detection module. The selection process is achieved by applying a learning approach using multiple additional features and a suitable metric in multi-feature space. The metric is derived by learning the underlying parameters using a small set of representative face samples. In this process the parameters are optimized by a Multiple Objective Optimization method based on the Pareto Archived Evolution Strategy. The learned metric in multi-feature space is then applied to a conventional classifier to filter out non faces from the first processing step. Support Vector Machines and K- Nearest Neighbours classifiers are used to test the performance of the optimized metric in multi-feature space. 1 Introduction Face detection is a well-known research area in image processing. It is fundamental in video surveillance and non-intrusive biometric applications where accurate detection is needed before recognition. Related research has come a long way since the survey made by Chellappa et al. one decade ago [1]. Within surveillance applications, a main challenge in face detection originates from the highly dynamic and non-rigid nature of human faces. Moreover, different face poses, facial expressions and occlusions make the problem very complex and in some cases unsolvable. The challenge is to detect human faces in complex surveillance environments regardless of their locations, poses, scales, resolutions and lighting conditions. There are mainly two different approaches to face detection in the literature. Feature-based techniques extract facial components such as eyes, nose and mouth as cues to deduce the existence of a face. Image-based approaches treats face detection as a classification problem embracing training and learning. In the last few years, image-based approaches have been the focus of research proving to be very effective in dealing with complex environments, varying illumination conditions and cluttered backgrounds. However, some of the best algorithms are still too complex for real-time processing as required in surveillance applications [2]. To implement an efficient automatic face recognition system, the techniques used to detect and locate faces in a scene have to be robust and quick. Colour is a powerful and fundamental cue that can be used to select potential candidate areas since colour segmentation is computationally fast and relatively robust to changes in scale, viewpoint and illumination conditions. However, to discriminate true face regions from other image areas additional low-level features are necessary. In this paper, a two-stage strategy for face detection in colour images is proposed. In the first stage a-priori knowledge about human skin colours is used to reduce the searching space of potential face candidates. Skin colours are approximated by a Gaussian mixture model in the YCbCr colour space. In the second step faces are detected among the candidate regions extracted with the skin colour model. The goal is to filter out false positives using additional primitives. This selection process is achieved by applying a learning approach using multiple features and a suitable metric in multi-feature space. The metric is derived by learning the underlying parameters using a small set of very representative face samples. In this process the parameters are optimized by a Multiple Objective Optimization method (MOO) [3] based on the Pareto Archived Evolution Strategy (PAES) [4]. The learned metric in multi-feature space is then applied to a conventional classifier to filter out non faces from the first processing step. Support Vector Machines [5] and K-Nearest Neighbourhood classifiers are used to test the performance of the introduced metric in multi-feature space. Selected results from computer simulations are also reported in this paper. 2 Skin Detection and Initial Face Extraction The method proposed in this paper consists two steps. The first step uses a conventional skin detection method to extract regions of potential faces from the image database. This skin detection step is based on a Gaussian mixture model in the YCbCr colour space. In the second step faces are detected among the candidate regions by filtering out false positives from the skin colour detection module. The objective of the first step: skin modelling is to find a decision rule that could discriminate between skin and non-skin pixels. The aim of this step is to compare and evaluate different skin colour models and thereby to estimate the properties of human skin colours. By doing so skin regions in colour images can be effectively detected and hence the search space of possible face candidates can be reduced. Skin-colour models exploit the fact that different human skin colour tends to form a tight cluster in colour spaces [1]. In the proposed approach the skin detection