Abstract |
Human face detection even though is performed instantly, effortlessly and with indicative accuracy by the human brain, for machine vision research is a matter still under development. Moreover, its wide range of applications like automatic data preparation for face recognition, content-based image retrieval or advanced human and computer interaction, make it a problem with both theoretical and practical values. The aim of this study was to introduce Convolutional Neural Networks as an efficient and fast face detector, able to operate in un-controlled image environments and without preprocessing. A convolutional neural topology is proposed, designed to be robust in varying image conditions and facial expression or other input deformations. The network was trained over a large enough training set of face patterns, coming directly from natural data, via the backpropagation algorithm. Using the trained convolutional filters of the network, a fast procedure for image scanning and face localization was devised, based purely on basic image processing operations. The system was tested in a series of large and difficult test sets exhibiting very high detection rates with a few and sporadic false alarms. The comparison with the current stateof- the-art systems in common benchmark sets revealed that the proposed system is the best performing general-purposed face detector of the reported literature. Furthermore, the tolerance of the network in a series of possible input deformations was measured and verified in conducted sensitivity analysis experiments.
|