Face Verification Algorithm Based On MLF-AWCN And The Fusion Of Diverse Differences Of Features

Recently,the study on the deep convolutional networks has taken face verification technology to new heights.On the basis of the achievement in the past,we design a new deep convolutional networks architecture for face verification.Based on general deep convolutional networks architecture,we propose the accumulative weighted convolutional networks with multi-level features melded(MLF-AWCN).This algorithm can avoid the vanishing gradient problem effectively through adding a softmax network over feature layer at each level.Meanwhile,weighting each softmax network improves the flexibility of feature learning.Then,features at all level are melded into a comprehensive feature whose dimension is decreased through linear projection.This process guarantees to extract abundant image information,thus improving the verification performance significantly.Data analysis shows that the feature components are uncorrelated and identically distributed approximately,which indicates the advantage and effectiveness of the feature learned through networks.We propose a new comparison method which is based on LDA and the fusion of diverse differences of features,aiming to cover the shortage of cosine.The cosine normalized by absolute value(cosAN)is designed to measure the difference of lengths between two vectors.So,the cosAN and the cosine are complementary.The comparison algorithm based on LDA and the fusion of diverse differences generates a new measurement for the differences of two vectors through melding the cosine,the cosAN and the lengths of two vectors.Experiments indicate the proposed algorithm is effective.Finally,the experiments show that,the proposed method achieves the verification rates of 99.9%,100%,98.8%,and 99.6%on four test sub-databases in FERET which are Fb,Fc,Dup1 and Dup2 respectively,when the false positive rate is 0.1%.