Face Detection and Eye Localization in Video by 3D Unconstrained Filter and Neural Network

Abstract

Frequency domain 3D-filter designing for automatic face detection and neural network based searching algorithm for eye localization of detected faces in video sequences is proposed. A series of spatiotemporal volumes are constructed from the video sequences of faces by concatenating the frames of a single complete cycle of face position is used to design a 3D unconstrained correlation filter by classical Fourier approach. The Unconstrained Optimal Trade-off Synthetic Discriminant Function (UOTSDF) filter is generalised here into a video filter of 3D spatio-temporal volume. After extracting the facial region by 3D correlation filter in frequency domain of the video frames, a neural network is employed to locate the eyes. The novelty of the face detection in video by frequency domain analysis and fast eye searching by parallel neural net of Generalised Regression Neural Network (GRNN) is validated with the benchmark database like VidTIMIT video database.