Face detection in historic documentaries with a cascaded classifier

Face detection aims at detecting and localizing an unknown number of faces in a still image or video frame. The challenges are to detect all faces while keeping the false positive rate small and to minimize the detection time per frame.
We study face detection in the context of historic documentaries. The source material for this work are films of the Soviet film maker Dziga Vertov that date back to the 1920's. The digitally available material bears major image deficiencies including flicker, scratches, dirt, bad lighting and contrast and visible frame lines. Naturally, the material is monochromatic and silent.
Based on a literature survey on different approaches for face detection, we select a method introduced by Viola and Jones for this investigation.
Their approach employs a cascaded classifier, i.e. a sequence of nodes, that distinguishes faces from non-faces. These nodes are organized as a hierarchy of classifiers that are built from simple, Haar-like features.
The main advantage of using a cascade is that only a moderate false-positive rate is needed for individual nodes as the individual rates multiply up to the overall false-positive rate.
We describe how the detection framework is set up for and adapted to the historic material and how it is implemented. Additionally, we suggest several post-processing steps to ameliorate the false-positive rate.
Finally, we provide detailed results for several sample scenes from the documentaries, and analyze the performance of the training and detection stages.