Ying Zhu

Fast face detection using discriminant wavelet features

 Abstract:

Face detection is a very useful technique with broad applications such as video surveillance, human computer interaction and content-based image/video retrieval. Detecting human faces in natural images is a 2-class pattern classification problem, where we need to distinguish the class of face patterns from all the "nonface" (background) patterns. Our work is focusing on developing fast detection methods since most of the state-of-art face detectors can hardly be useful in real-world applications due to their high computational complexity. Two techniques are used to speed up the detection: discriminant wavelet features and sequential Bayesian detection. This allows us to quickly rule out background regions with a small feature set. I will talk about how we evaluate the discriminability of wavelet coefficients and choose the most efficient feature set in terms of class separation accordingly. Some experimental results will also been shown.