Abstract:

A neural network-based face detection system was proposed.Unlike similar systems which were limited to detecting upright frontal faces,faces were detected at any degree of rotation in the image plane.Multiple networks were employed;a "router" network first processed each input window to determine its orientation and then used this information to prepare the window for one or more "detector" networks.The training methods for both types of networks were presented.Sensitivity analysis on the networks were performed and empirical results were presented on a large test set.Finally,preliminary results were presented for detecting faces rotated out of the image plane, such as profiles and semi-profiles. Experimental results show that this system has better performance in accuracy and detection speed,and it is faster than a single network system.

Key words: rotation face detection, neural network ensemble, router network, network window