The Establishment Of A Face Recognition System Based On Deep Learning And An East Asian Face Database

Benefitting frorm the development of deep learning and the establishment of large-scale face database,it has been making a significant breakthrough in the area of face recognition.However,these large-scale face databases majorly consisted of Western faces,and the existing East Asian face databases are not large enough for the training of deep neural networks.The face recognition systems trained by these databases will not obtain desired results to deal with East Asian face.Our work aims to establish a large-scale East Asian face database to fill vacancies in this area while designing a face recognition system for East Asians.The major results of our work include:a small,East Asian face database which is fully calibrated by hand,including a total of 193,602 images of 4836 people;establishment of a more than 3 million pictures,8,000 individuals of the East Asian face database(to our knowledge,there is no public large-scale East Asian face database,and the only public database consists of millions of pictures is MegaFace);a face recognition system for East Asians is designed,which uses the deep learning method to extract the image features,then uses the machine learning method to reduce the dimension,and finally a Joint Bayesian model is trained for the face recognition.The experimental results show that this method is superior to the usual face recognition systems(trained from the Western faces)in the recognition task of East Asian face picture.We randomly selects 10,000 pairs of East Asian face data,and it can achieve 97.92%recognition accuracy.A new method based on RANSAC is introduced to realize the automatic calibration of face picture and update the joint Bayesian model at the same time,which provides a convenient scheme for automatic calibration of large-scale face database.We also designs a modified triplet-loss function,which avoids the negative sample selection problem(selecting the semi-hard ones)of traditional triplet-loss and improves the training efficiency.