2. Face Detection in Complex Background ( Online Demo )

We used our information-based learning technique in the the context of face detection. We trained an 11x11-pixel model with image examples of faces and backgrounds. Examples of faces were obtained from a subset of the frontal-view images of the FERET database. A collection of images of a wide variety of scenes with no frontal-view faces were used as negative examples. Grey level images are re-quantized to four intensity levels, so that observation vectors consist of 121=11x11 pixels with 4 possible values each.

Scale-invariant face detection is carried out via multi-scale search using the likelihood ratio model obtained with the learning technique. We first compute the likelihood ratio of each sub-window of the multi-scale pyramid of images. Then, face candidates are obtained by selecting the position with local maximums. Heuristic rules are later used for candidate validation combining the candidates from different scales and neighbor regions.

Using a similar approach, this work is extended for the detection of the eye corners. A combination of the eye detection confidence level with that of the overall face detection is used for improved face candidate validation.