| Face recognition is one of the most important research topics in artificial intelligence,computer vision and pattern recognition.It has attracted significant attention from both industry and academia due to its advantages as an effective biometric feature,such as being portable and difficult to steal.With the popularization of intelligent devices,the products equipped with the face recognition technology are widespread,e.g.,the face recognition based unlocking application in the smart phone.Due to its wide range of applications,face recognition has created important economic and scientific value for the human society.However,because of the internal and external factors,there are many problems that need to be solved in the face recognition tech-nology.On the other hand,with the rapid development of computational capacity,deep learning shows its powerful ability for representation learning in computer vi-sion.Therefore,it's worthy to solve the existing problems in face recognition with deep learning.The main works of the thesis are introduced as follows.Firstly,we introduce the basic concepts of deep convolution neural networks(DNNs)and face recognition methodology in detail.We divide the face recognition algo-rithms into four major categories according to its development history.Then,the representative algorithms of each category are introduced.Secondly,we investigate and analyze the existing DNNs based face recognition al-gorithms in depth.Then,two popular DNNs training frameworks for face recog-nition are summarized,i.e.,classification based and metric learning based training framework.And we propose an effective training framework for DNNs based face recognition by taking advantage of the two existing frameworks.Moreover,we intro-duce the adaptive margin,instead of a fixed one,into the training framework in this thesis.The adaptive margin is beneficial for generating high-quality training sam-ples.Experimental results show that the proposed framework obtains a reasonable trade-off between performance and training time.We verify the effectiveness of the proposed DNNs training framework on the face verification and identification tasks.Finally,we discuss the robust face recognition under large noises.After analyzing the existing solutions,we divide the face recognition under noises into two classes,i.e.,image denoising preprocessing based and robust face representation based meth-ods.The image denoising preprocessing methods may not effectively improve the performance for the subsequent recognition task.But the robust face representation directly fits the distribution of the noisy face image,resulting in a slightly better recognition performance.Based on the analyses,we introduce the ideas of residual learning and multi-task learning to robust face recognition under noises in this thesis,where a noise resistant DNNs is proposed.The effectiveness of the proposed method is verified on the multiple face datasets on the face verification and identification tasks. |
