Research On Face Recognition Based On Improved Convolutional Neural Network And Its Application In Bank Vault Monitoring System

In biometric recognition technology,face recognition technology is its classic example,which has a good application prospect in real life.As a necessary research technology of many scholars,face recognition technology has many research results.At present,face recognition can achieve more accurate recognition efficiency in simple background or fixed scene.However,accurate face recognition in complex scene is still a challenging technical topic.In the process of face detection and feature extraction,the background,light,face posture,motion occlusion and expression can affect the above processes,resulting in poor recognition stability and reduced algorithm recognition efficiency.Therefore,in the process of identity verification in the face of the banking financial system,the face detection and feature extraction algorithm with high robustness and efficiency is very important for face recognition technology Important practical application value.I work in the bank.The vault is an important place for the bank to store cash and other assets.The safety of the vault plays a very important role for the bank.The establishment of a modern vault system can greatly improve the safety of the vault.Since the vault can only be accessed by the relevant personnel of the bank and the personnel of the security company,no more than 100 people are qualified to access the vault,therefore,I have designed a product with moderate storage space and try not to use too complex network model to ensure high stability,high authentication speed and high accuracy.Based on the minimum resources / cost,I can achieve the highest accuracy of face recognition.First of all,this paper expounds the background and significance of the subject,combs the current domestic and foreign research literature on face recognition technology,describes the three main processes of face recognition technology in this stage,as well as the relevant models of machine learning algorithm and depth learning algorithm applied in each step,summarizes and discusses the problems of the current mainstream algorithm.Secondly,in the face of the problem that the accuracy of face recognition will be reduced due to many factors such as human posture,facial expression and motionocclusion,a face recognition network model based on improved residual neural network is proposed.Under the condition of deleting the number of residual units and reasonably increasing the width of residual neural network,the degree of residual neural network model is clearly given.The results show that the improved residual neural network model has obvious recognition accuracy compared with the original neural network.And thirdly,the two-dimensional Gabor filter is added to the convolution neural network structure,and the convolution neural network(GCNN)is constructed on five scales respectively.The results of direct classification with GCNN,feature extraction with GCNN and SVM classification,and Gabor feature extraction with PCA and SVM classification are compared and analyzed.The experimental results show that the effectiveness of GCNN,whether it is a direct classification or a feature extraction and then sent to SVM for classification,its training process is quite stable.Finally,in order to test the accuracy of the system and the accuracy of face recognition in the background of bank financial monitoring,the monitoring system is tested in practical application.Through the statistics of the test results and functional data,the specific experimental results are: 1)when the facial expression is exaggerated and the posture is close to the face front picture,the system has a high acceptance rate.The tilt angle of the face is not excessive.Even if there is a top view angle,it can be correctly recognized,so that when there are hand movements on both sides of the face,it can also be correctly accepted and recognized.In the sample images that can not be effectively recognized,the face is occluded by glasses,which leads to recognition errors or system unacceptability.2)The bank monitoring system can recognize about 30 face images within 1 second,so the system can realize the application in normal scenarios,and also meet the real-time requirements.