Design And Implementation Of GPU Based Large-scale Face Recognition System

During recent years,great progress has been made in deep learning based face recognition,which is one of the most active research field in computer vision.Face recognition has been widely used in public security,finance and many other fields.Although the accuracy of face recognition has reached in a satisfactory manaer under controlled conditions,the performance in real application remain a challenge,because of the effects of illumination,pose and disguise.The increasing of face data scale may lead to low accuracy and high query delay in face search.In this thesis,the basic theory and key technologies of face recognition were systematically investigated for solving accuracy and query delay problems brought up by illumination,occlusion and pose impact and massive face data,the complex background of face recognition and large-scale face search problems were explored in depth.In addition,a face recognition model and a fast face search method were proposed.The main contributions of this thesis are summarised as follows:(1)A face recognition algorithm based on convolutional neural network(CNN)was proposed.The algorithm uses less number of kernels,introduces the residual units among convolution layers,combines softmax loss and center loss for network training.We trained the CNN model on MsCeleb-1M data set and evaluated it on LFW and YTF.The experiment results show that the trained model achieves 99.65% accuracy in LFW and 96.5% accuracy in YTF.At the same time,the makespan of feature extraction on a single Titan GPU was 6ms per face image.(2)Based on the aforementioned face recognition algorithm,we proposed a batch-oriented method for face feature extraction.When the batch size was set to 32,the average speed of feature extraction was 1.4ms per face image.(3)We also proposed a fast face search algorithm which is based on two level index.The search procedure is divided into two separate phases: multi inverted index filtering and float vector extract re-ranking.By leverage the power of GPU,our implementation is able to search 5 million face data within 10 ms.(4)A GPU-based face recognition scheme was proposed.It uses master-slave architecture for supporing the real-time retrieval on large-scale face data,which is able to serve a searching request quickly,efficiently and accurately.We implemented the face recognition system and evaluated it,the experiment results showed it can provide real-time,high-performance face recognition services.