ROBUST FACE DETECTION FOR MOVING PICTURES UNDER POSE, ROTATION, ILLUMINATION AND OCCLUSION CHANGES

Abstract

Face detection has been a key step in face analysis systems for decades. However, it is still a challenging task due to the variation in image background, view, pose, occlusion, etc. This paper proposes a simple and effective tool to detect human faces in moving pictures under such conditions. An improved approach aiming to reduce impacts of illumination, scale and connection of faces to receive rapidly skin homogeneous regions considered as the most potential face candidates is presented. A hybrid classifier, applied in retrieved face candidates, is based on template matching and appearance-based method providing a robust face detection. This verification achieves advantages of the powerful discrimination of Local Binary Patterns (LBPs) and the high speed detection capability of embedded Hidden Markov Models (eHMMs). Experiments were performed with different image databases and video sequences such as NRC-IIT facial video database, Caltech database, etc. Our system is effective in detecting not only frontal faces but also profile, rotated, occluded and connected ones for real-time application.