The Research On Key Technologies Of Face Verification System Based On Second-generation Identity Card

Face verification combines several hot subjects together including computer vision,machine learning,pattern recognition etc.It has a wide range of applications in a number of areas,such as public security,finance and military sectors.Recently,with the rapidly popularizing of AI and related technologies' industrialization,some light applications of face verification based on Chinese second-generation identity card shows up in banks,airports,train stations,etc.,aiming to achieve auto re-identification.A complete face verification system based on ID card image consists of face detection and alignment,image preprocessing and face verification.The entire face verification process should perform circular face comparison on the scanned image of ID card,the image read by particular card reader,and the real-time image collected by camera.Although face recognition has been deeply researched in the past twenty years,it's still a challenging task due to the combination of complex imaging conditions,age progression,low resolution,illumination change,random noise,etc.More specifically,the image stored in the ID chip is of extreme low resolution and impossible to be processed directly.These variations result in low true-positive rate and high false-positive rate of the system,which is far from satisfactory for the security needs.We have done researches on the key technologies of face verification based on second-generation ID card.Firstly,the super-resolution reconstruction of the chip image is achieved by reusing the particular compression procedure and then utilizing very deep CNN super-resolution model(VDSR)to learn the mapping function.Secondly,we propose an algorithm based on multi-scale bilinear pooling CNN model for across-age face verification.We use CNN to learn features' robustness to age changes end-to-end,and bilinear pooling model is introduced into the network for the second-order pooling over multi-level features to enhance the ability of discrimination.Finally,we use center-loss as the face verification supervision signal to achieve better clustering results by constraining the similarity of features from the same label.We evaluate our model using the second-generation ID card images we collected before.Three experiments are conducted to compare faces between chip images and camera images,scanned images and camera images,chip images and scanned images.Experiments show that the multi-scale bilinear pooling convolutional neural network with the chip's super-resolution preprocessing achieves accuracy of 70.68%,65.50%and 96.97%respectively at a 0.1%false positive rate,which exceeds the open source face recognition engine Seetaface.