Research On Deep Learning Based Face Recognition Under Complex Illumination Conditions

Deep neural network has been widely studied in computer vision area for its standout performance.Compared with traditional methods with hand-crafted features,convolutional neural network can learn better face representations from large amount of data through the end-to-end architecture.But the performance of deep neural network models under complex illumination conditions has not been thoroughly researched.In this paper,we study face recognition technology based on deep neural network under complex illumination conditions.The main contents of this paper are as follows:Firstly,we build a large dataset of faces under complex illumination conditions to study the effect of illumination on face recognition.330000 face images of drivers are collected in the dataset,and the dataset contains strong light,weak light,locally overexpose and other illumination problems.Secondly,we analyze the performance of different neural network structures on the dataset.The performance of three convolutional neural networks(LeNet?AlexNet and VggNet)is studied and proved to be effective for face recognition under complex illumination conditions in this thesis.Thirdly,we study data augmentation technology.Adding illumination transformations on training dataset by converting images with illumination based image processing methods.The neural network trained on augmented dataset can learn illumination invariant features and is robust to illumination problems.Finally,we study ensemble batch normalization to tackle the problem of the training time and instability of deep neural networks.This method will make the neural network converge more stably and fast without degradation in performance.This paper discusses the existing problems of deep neural networks under complex illumination conditions,and provides guidance for designing and optimizing on neural networks for further researches.