The Construction Of Large-scale Asian Face Data Set

In recent years,with the vigorous development of deep learning,more and more researchers and Internet teams are thrown themselves into this area.The proposed of Convolutional Neural Network has a huge impact on the field of computer vision.At the use of deep learning method,the accuracy of face recognition has got breakthrough progress,and gradually become a very practical way.The promotion of it mainly relys on two aspects:optimizing the network models and enhancing the training data sets.However,the public face data sets which are used for training depth of network are very scarce at present,especially for Asian face data set.Most sets are based on Europe and the United States face images.As is known to all,compared with east faces western faces have higher brow,deeper eye socket and higher cheekbones.The network model based on WebFace has an accuracy rate over 99%in LFW test data set.But according to our experiment,the model accuracy is significantly reduced when test the Asian faces,therefore,building a set of data used for training Asian face is very necessary and meaningful.To achieve this goal,we need to research the method of seting up a data set,upgrading building efficiency,reducing construction costs,the work involved including:a.Realized the building face the full processing of dataset,we can automatically calculate complete facial and manual intervention the combination of data acquisition,data processing of automated and manual review.In addition,given the scale needed for deep learning data set is large and has high requirements for quality.Face image data acquisition and manual review,the two larger workloads and costs in the process,developed for the corresponding application management systems.Face image of the Internet crawling and later manual review call out task management,data acquisition and labeling audits more efficiency.Automatic data processing stages,this study mechanism for achieving scores of image data,automatic tagging and filtering to achieve large amounts of image data,which greatly reduces the manual effort,enhance the efficiency of construction of data set.b.Wi th the use of above system and processing method,we construct a data set containing more than 5,000 Asian stars and a 500,000 pictures in image size.The data set covers gender-age Asian stars,and each facial image contained within the stars try to maintain a differentiated to facilitate web model of training.Experiments on network model in the same structure,face the depth of training data set based on Asian net work model WebFace(Europe and American face mainly)depth of the training model on face recognition results,and good results have been achieved.In addition,we also use web development simple application of face recognition systems,reflecting the depth of the network model of practicality.In short,this paper builds a complete construction of training data set face recognition system and actually creates a high quality Asian face massive data set,which reflects the effectiveness of the system.Experimental results show that Asian stars based on this data set the depth of the training network in identifying Asian faces,by contrast former European star data sets,has obvious advantages in accuracy.