Research And Implementation Of Face Recognition Algorithm Based On Deep Learning Under Unconstrained Conditions

Nowadays the face recognition technology has been widely used in many real-world scenarios.In fact,the face recognition rate is still relatively low in many unrestricted conditions(including light,angle,occlusion,makeup,etc.),which limits the further development of face recognition technology.To this end,this thesis studies the face recognition technology under occlusion,and proposes a face segmentation algorithm based on semantic segmentation and a face completion algorithm based on occlusion area localization,and implements an occluded face recognition system.The specific research contents are as follows:1.This thesis proposes an automatic localization algorithm for face occlusion area.At present,the occlusion area automatic localization algorithm based on face semantic segmentation is only for rectangular mask design,and does not consider the case where the occlusion area has different shapes and textures.In view of this,this thesis proposes a semantic segmentation algorithm based on pyramid pooling model to realize automatic positioning of face occlusion area,and trains and verifies the MMCCH dataset collected in this thesis.The experiment shows that the method has higher segmentation accuracy.2.In this thesis,a face completion model based on occlusion localization is proposed.Aiming at the problem that the current face completion model is too free,an improvement based on the VAE-GAN model is proposed,which combines the random distribution features and the face Gaussian distribution features in VAE as the input of GAN,so that the model become more controllable and generate realistic face completion results.In addition,in view of the fact that the current face completion model relies too much on the rectangular occlusion area,this thesis improves the model by partial convolution,so that the completed occlusion area pays more attention to the information of surrounding valid pixels,and can deal with irregular masks with different positions.In this thesis,seven kinds of masks are set in the test set of celebA.The results of the experiments show the improved model is effective.3.An occlusion face recognition system based on face occlusion area localization algorithm and face completion model is implemented.The model firstly detects the occlusion area and then completes it.Then the FaceNet face recognition model is used to extract the 128-dimensional face features and then face-face comparison with the face database feature.Finally,the top three most similar faces are shown.In this thesis,the verification experiment of occluded face recognition system is carried out on the MMCCH dataset.The experiment proves that the face recognition rate is improved significantly after the face completion when the occlusion rate is lower than 25%.