Research Of Face Recognition Based On Convolutional Neural Network

Face recognition is an important branch of pattern recognition and computer vision.Face recognition has been widely used in scenes such as security monitoring,identity recognition,human flow analysis,image reconstruction,and human-computer interaction,which has attracted more and more attention.Face recognition process includes image preprocessing,key point detection,cropping and alignment,feature extraction,identification,and verification.Face recognition,as a biometrics-based recognition technology,has a high discriminability.It is a non-mandatory,non-contact,high accuracy,easy to collect,high concurrency,non-invasive biometrics technology.As deep convolutional neural networks continue to improve in structure and number of layers,face recognition has made tremendous progress.The core task of face recognition is to obtain features with sufficient discriminative power.In a typical face recognition process,the deep convolution neural network trained by Softmax loss function to extract face features.Usually,the output of the penultimate layer is used as the face feature.Given a pair of face images,the cosine similarity or Euclidean similarity between face features is calculated,and the threshold is calculated to distinguish the similarity between them.The Softmax loss function of the traditional deep convolutional neural network performs well in the classification task.However,for a million-level face dataset,it lacks sufficient discriminant ability,and does not explicitly encourage learning discriminative features,and cannot obtain optimal results.In order to solve this problem,this paper proposes an improved loss function,which can reasonably supervise the training of deep convolution neural network to obtain the discriminant face features with compact intra-class and inter-class separation.The main work and innovations of this paper are as follows:The traditional Softmax function has insufficient separability,while the new Softmax function expands the separability by parameters and has stronger class separability.The traditional center loss function maintains the class center for each category,and punishes the distance from the same category to the class center.Only considering the intra-class distance penalty,ignoring the inter-class distance penalty,and The new central loss function not only punishes the traditional intra-class distance,but also punishes the inter-class distance,so that the network can simultaneously obtain the characteristics of compact intra-class and separation inter-class.Therefore,this paper combines two new loss functions to train deep convolutional neural networks on large-scale datasets and analyze the loss function curves.This paper first verifies the effectiveness of the improved loss function on the MNIST handwritten dataset and visualizes the results.The Inception-ResNet network combines the residual network and the Inception network,while expanding the width and depth of the network,preventing gradient explosions and disappearances,and incorporating information from different local receptive fields.Therefore,this paper trains Inception-ResNet network on CASIA-WebFace to verify the effectiveness of the improved loss function on LFW.Experiments show that the improved algorithm can reasonably supervise a single deep neural network on large-scale data sets,and obtain the features of separation between classes and compactness within classes,and obtain competitive results compared with other algorithms.