Estimating Conformity of Head Yaw to the ICAO Standard using a Convolutional Neural Network

摘要

Face images are widely used in authentication documents to recognize individuals. Face quality assessment can be used to enhance the efficiency of face recognition algorithms. One of the most important challenges in the design of face image quality algorithms is the measurement of parameters such as darkness, light conditions, head position and facial expressions. In recent years, traditional identity documents have been replaced by electronic documents with biometrics traits. International Civil Aviation Organization (ICAO) has imposed some requirements for face image features, according to them any image used for biometric applications should be conform to these requirements. In this research, a model has been developed to measure the match between orientation of the head rotation around the vertical axis (Yaw), which is one of the requirements of ICAO (ISO/IEC19794-11). The basis of our work is the use of convolutional algorithms and supervised learning in the field of deep learning methods. Using the reconfiguration of VGG-Face deep neural network and a combination of different databases, we have trained the deep learning model. Then, to compare the proposed method with the results of other articles, we tested the learned network on the PIE database and accurately identified 96.5% of the correct diagnosis. Finally, in order to test the genericity of the proposed method, we tested it on the CSIE Robotic database, where we obtained an accuracy of 97.8%.