Research On Face Recognition Under Complex Illumination Based On Deep Learning

Face recognition,as a biometric technology with strong security,fast recognition speed and high accuracy,is one of the important research directions in the field of artificial intelligence.With the rapid development of big data and deep learning,the performance of face recognition has become more and more close to human eye recognition,and face recognition technology has been widely used in police security,public security,social security and other fields.However,in an unlimited scene,complex lighting,pose transformation,age and expression changes,physical occlusion,etc.have a great impact on the face image.Especially under complex lighting conditions,the face image will appear serious distortion and lose a lot of effective information,resulting in a rapid decline in the accuracy of face recognition.In view of the above problems,the main research contents of this thesis are as follows:(1)Aiming at the problem of poor quality and large distortion of face image under complex illumination,the low-light enhancement model of Generative Adversarial Network architecture is optimized,and the illumination adaptive face recognition method based on Generative Adversarial Network is proposed from two aspects of image preprocessing and generating image feature constraints.First,in order to enable the original low-light enhancement model to process strong light images adaptively,a illumination adaptive threshold differentiator module is designed to split the face images of strong light,dark light and normal light,and the strong light face is transformed into negative image domain as the input of the original network model,so as to enhance the generalization processing ability of the original model to the light face image;Secondly,in order to make the feature specificity of the face image generated by the model stronger and the facial features clearer,a face analysis module is introduced,and the attention map is constrained by the face mask map,and the feature loss is constrained by the face feature extraction network.Validation experiments were conducted on the Ex-YaleB dataset,CAS-PEAL-R1 dataset,and AR dataset containing faces with varying lighting conditions.The accuracy of face recognition reached 82.1%,70.2%,and 72.5%,respectively.The experimental results showed that this method can effectively handle complex lighting faces and improve face recognition rate.(2)The recognition accuracy of face images generated using the above methods for face recognition remains to be improved.Therefore,a complex illumination face recognition method based on attention feature fusion mechanism is proposed.This method studies attention mechanism and feature extraction network,and proposes an attention feature fusion network structure based on residual network using global local channel attention mechanism.It improves the network feature extraction ability by integrating global attention and local context attention;In addition,using the maximum characteristic graph activation function to replace the Re LU activation function in the original network structure can further improve the performance of the network model with sufficient iterations.Through ablation experiments and comparative experiments,the accuracy of facial recognition on the Ex-YaleB dataset,CAS-PEAL-R1 dataset,and AR dataset containing face changes in lighting was verified to reach 83.4%,71.3%,and73.9%,respectively.The experimental results show that this method can further improve the recognition rate of network human faces.