Research On Face Information Processing And Acceleration Based On Deep Learning

Face recognition algorithms have been widely used in public security,national economy,and social service related fields.Although the current face recognition algorithms based on deep learning have achieved very good performance on public datasets,their recognition capabilities are still insufficient in complex scenes,such as changing lighting,angles,and distances;some algorithms are often complex and difficult to achieve improved performance.They require a large volume of storage space to store a large number of samples for training.They are limited by network bandwidth,latency,and data storage issues,making the efficiency and performance of local training and inference suffer.Therefore,this thesis mainly studies the optimization and acceleration technology of face information processing algorithm based on deep learning.The specific work includes the following aspects:This thesis studies a face recognition algorithms in complex scenes.In view of the factors such as lighting,posture,distance and expression in complex scenes that affect the convergence performance of deep learning algorithms,we design a loss function based on angle adaptation,which can be dynamically adjusted during the training process for difficult samples in complex scenes.The loss in the model training process improves the convergence performance of the deep network.Extensive experiments on related data sets,verified the performance of the proposed method by comparing with the current mainstream loss function methods.This thesis studies a face attribute recognition algorithm in complex scenes.Face attribute recognition algorithms are more sensitive to the influence of scene factors than identity recognition algorithms.Therefore,in view of the large deviation of face pose in complex scenes,complex lighting conditions,diverse backgrounds,and occlusions in some pictures,this thesis designs a deep neural network for face attribute recognition based on multi-scale features,through effective feature extraction of different scales,enhances the deep neural network's ability to extract face attribute features,and can effectively combat the impact factors of complex scenes.The validity of the method proposed in this thesis is verified by comparison experiments on related data sets with the current mainstream face attribute recognition networks.This thesis studies a training acceleration scheme for local deployment of face recognition algorithm.At present,mainstream face recognition algorithms based on deep learning often use distributed training methods to relieve the pressure of data storage and calculation on a local single server,and distributed training often faces the problem of data transmission bottlenecks.To address such a bottleneck problem,this thesis designs and implements a data pre-access method and a local caching scheme.Through a large number of load tests on the experimental platform,the effectiveness of the proposed acceleration scheme is verified.This thesis studies an acceleration scheme to accelerate the inference of locally deployed face recognition algorithm.face recognition algorithm.In view of the large load and high instantaneous concurrency in local deployment scenarios,this thesis applies related model pruning to effectively reduce the amount of model computation,and designs a an acceleration scheme that effectively balances a large number of local loads.The acceleration scheme realizes load balancing of multiple locally deployed parallel computing servers,and further uses multi-process technology to accelerate parallel processing on a single computing server,which can effectively handle the local high concurrent computing pressure.A large number of experiments on the built hardware experimental platform have confirmed that the scheme proposed in this thesis is effective and robust.