A Novel Coupled Template for Face Recognition Based on a Convolutional Neutral Network

Abstract

Skin segmentation can effectively improve accuracy of face searching in a picture. However, it is a difficult problem to segment face skin from a photo with complex background. In this paper, a novel coupled template for face regions extraction after skin segmentation is proposed to overcome the difficulty that face regions are largely sticky to similar skin backgrounds. The algorithm based on the coupled template is able to separate the face regions from similar skin regions correctly and enhance the correct rate of face region detection largely. Moreover, a prior knowledge of standard face region can be embedded into face searching process to locate the face position and then a well-trained convolutional neutral network is used to recognize faces so that the accuracy of face recognition can be improved further. The novel approach has a good adaptability to the image with complex background that results in many large sticky similar skin blocks. The classical basic architecture of convolutional neural network LeNet-5 is employed and only focuses on the accurate located-areas. The high recognition rate and low missing detection are obtained for the pictures with complex background especially with a large number of color blocks similar to skin.