Research On Face Recognition Based On Convolutional Neural Network

Face recognition is one of the biometric technologies,which has been widely used in many areas of our daily life,including the access control system for residential home security,the attendance system for company management,the identity authentication system for citizen’s travel,the payment system for online shopping,the public security system for criminal investigation,etc..Although fac e recognition has the advantages of easy operation,initiutive results,and good concealment,due to the randomness of the shooting conditions,the collected face images are often affected by the lighting,the expression,and the occlusion.In this thesis,a new algorithm is proposed which integrates the traditional face feature extraction methods with the convolutional neural networks to solve the above issues and the following research work had been performed:(1)Compared and analyzed HOG 、 LBP feature extraction algoirthms,and studied the working principle of the convolutional neural network and the CBAM attention mechanism;(2)A face recognition model based on Le Net-5 network is developed in addressing the changes of face expression and the existence of the occluders.The model adds the CBAM attention mechanism after the pooling layer so that the network model focuses on the feature parts that have a major impact on the classification results during the training process.At the same time,the image features extracted by the network through the attention mechanism are weighted and fused to obtain the final face features.In addtion,a regularization method is added to the network to prevent overfitting,and a momentum term is added to the Adam optimizer for parameter optimization to speed up the convergence of the model.The experimental results on the ORL and CASIA-Face V5 datasets show that the recognition accuracy of face images by the above-mentioned improved model is improved by 2.0% and 2.69%;(3)In addressing the issue of illumination changes in face images,an algorithm is proposed which integrates the improved LBP operator(ULBP)with the vonvolutional neural network.The proposed algorithm first extracts the LBP features of the face images,and uses PCA to reduce the dimensions of the extracted face features,and then trains the convolutional neural network with the original face images.The relevant experimental results show that after adding this improvement,the recognition accuracy on the two datasets is increased by 0.75% and 0.58%,which demonstrate the feasibility and effectiveness of the algorithm.