Research Of Multi-Pose Face Recognition Based On Deep Binary Convolutional Neural Network

Multi-pose face recognition plays an important role in computer vision,the most common research methods are 2D method and 3D method.2D methods include extraction of invariant features,reconstruction of virtual views and regression model.3D methods include point-based modeling and pixel-based modeling.In view of the defects in 2D methods,this paper aims to combines traditional features and deep features to extract edge gradient with direction and scale-rotation invariant information of face image for face pose recognition.This paper first proposes a binary convolutional neural network of enhanced edge gradient directional information.Secondly,a scale-rotation invariant deep binary strategy is proposed for multi-pose face recognition.The specific work is as follows:1.In order to enhanced the feature robust of multi-pose face recognition,an enhanced edge gradient binary convolutional neural network is designed for recognition.Firstly,this paper proposed ROILBC(Region of Interest Local Binary Convolution)to select Region of Interest in face pose picture.Next,the DR-MGPC(Dimensionality Reduced Modified Gradient Pattern Convolution)is proposed to optimize features,then,the Enhanced DR-LDPC(Enhanced Dimensionality Reduced Local Directional Pattern Convolution)is proposed to extract the enhanced edge gradient direction features.Finally,Adopted Histogram similarity,chi-square distance,and Bhattacharyya distance as judgments to discriminate different face pose.Extensive experiments conduct on FERET and CAS-PEAL-R1 databases,the experimental results show that our method significantly outperforms other approaches on accuracy and computational efficiency.2.In order to solve the problem of scale-rotation transformation in multi-pose face recognition,the framework based on scale and rotation invariant deep binary features is designed in this paper.Firstly,this paper proposed ASRMs(Active Scaling and Rotating Masks),on this basis,we establish DOBCP(Difference of Binary Convolution Pyramid).Then,the extremum points located and eliminate distractions in DOBCP,furthermore,we endowing the feature maps in DOBCP powerful of scale and rotating invariant and construct description block.Finally,the feature descriptor is established on feature points of description block and conduct multi-pose face recognition.Extensive experiments conduct on FERET and CAS-PEAL-R1 databases,the experimental results show that our method significantly outperforms other approaches on accuracy.