Design And Implementation Of Face Recognition System Based On Convolution Neural Network

Convolutional neural networks(CNNs)are at the core of most state-of-the-art computer vision for a wide variety of task.It mainly benefits from the large-scale training data and the end-to-end learing framework.However,it is hard to look into the network and figure out what it has learnt.So I use some deep learning frameworks to validate the ability of convolution neural networks to fit nonlinear functions and visualize some classic convolutional neural network models to help understand what each neuron has learned and thus what computation it is performing.Since 2014,the quality of network architectures significantly improved by utilizing deeper and wider networks.The depth of representations is of central importance for face recognition.The number of stacked layers can enrich the “levels” of features.As more layers are added to a suitably deep model leads to higher training error because bigger size means more parameters,which makes back-propagation slower to converge and prone to overfitting.This paper introduces several different ways to improve the depth of the convolution neural network and designs new network strctures based on these methods.In most of the available CNNs,the softmax loss function is used as the supervision signal to train the deep model.For face recognition task,the deeply learned features are required to be discriminative and generalized enough for identifying new unseen classes without label predicition.Howerver,the softmax loss only encourage the separability of features,the resulting features are not sufficiently effective for face recognition.As alternative approaches,triplet loss and center loss was proposed.Both can enhance the discriminative power of the deeply learned features highly.In this paper,we compare the influence of these two loss functions on the performance of CNN in face recognition.At last,we get a neural network consisting of resnet and center loss and use the CASIA-Webface dataset for training.On the widely used Labeled Faces in the Wild dataset,it achieves 98.46%.Finally,we implement the face recognition system based on the depth neural network,including the establishment of the database and the face recognition,and achieved good experimental results.