Face Detection in Complex Natural Scenes

Face detection is an important preliminary process for all other tasks with faces, such as expression analysis and person identification. It is also known to be rapid and

automatic, which indicates that detection might utilise low-level visual information. It has been suggested that this consist of a ‘skin-coloured, face-shaped template’, while

internal facial features, such as the eyes, nose and mouth might also help to optimise performance. To explore these ideas directly, this thesis first examined how shape and

features are integrated into a detection template (Chapter 2). For this purpose, face content was isolated into three ranges of spatial frequency, comprising low (LSF),

mid (MSF) and high (HSF) frequencies. Detection performance in these conditions was always compared with an original condition, which displayed unfiltered images in the full range of spatial frequency. Across five behavioural and eye-tracking experiments, detection was best for the original condition, followed by MSF, LSF and HSF faces. LSF faces, which provide only crude visual detail (i.e. gross colour shape), were detected as quickly as MSF faces but less accurate. In addition, LSF faces showed a clear advantage over HSF, which contains fine visual information (i.e. detailed lines of the eyes, nose, and mouth), in terms of detection speed and accuracy. These findings indicate that face detection is driven by simple information, such as the saliency of colour and shape, which supports the notion of a skin-coloured faceshape template. However, the fast and more accurate performance for faces in the full and mid-spatial frequencies also indicates that facial features contribute to optimize detection.

In Chapter 3, three further eye-tracking experiments are reported, which explore further whether the height-to-width ratio of a coloured-shape template might be important for detection. Performance was best when faces’ natural height-to-width ratios were preserved compared to vertically and horizontally stretched faces. This indicates that this is an important element of the cognitive template for face template. The results also highlight that face detection differs from face recognition, which tolerates the same type of geometric disruption. Based on the results of Chapter 2 and 3, a model of face detection is proposed in Chapter 4. In this model, colour face-shape and features drive detection in parallel, but not necessarily at equal speed, in a “horse race”. Accordingly, rapid detection is normally driven by salient colour and shape cues that preserve the height-to-width ratio of faces, but finer visual detail from features can facilitate this process when further information is needed.