The Research Of Feature Representation And Face Recognition Algorithm Combining With Facial Attribute Based On Deep Learning

Deep learning,represented by convolutional neural networks(CNNs),is one of the most important new-emerging technologies and has been widely ap-plied in the fields of pattern recognition,especially computer vision,consid-ering its simplicity in distribution,convenience in training and superiority in performance.In the field of face recognition,training CNNs with numerous images for feature extraction to generate features has almost been regarded as a standard procedure.As the most significant visual characteristic for human faces,face attributes literally have strong correlation with recognition of faces'identity.Meanwhile,a face recongnition model in good performance is required to be robust to those external face attribute.So it makes a good sense for further resear-ch to explore the impact that face attributes exert on the face recongition task as well as full utilization of attributes' information to promote the accuracy of recognition in the era of deep learning.In this work,the author has carried out a detailed exploration on the rela-tionship of face attribute and face recognition in two different aspects.In the first aspect,the author goes straight through the information of attribute labels and proposes a multi-task learning fr-amework of jointly training the attribute recognition and identity recongition.The author also designs reasonable loss functions to ease the data imbalance in the category distribution of attribute la-bels as well as to get better face feature representation.Experiments show that the results of multi-task learning reveal a significant promotion compared with those single-task learning methods under the same setting in both attribute and identity recognition tasks.In the meantime,the results of the work is compara-ble to the state-of-the-art results in public test dataset,i.e.,LFW and MegaFace.In the second aspect,the author quests on the import of face attribute informa-tion from the perspective of training procedure.As commonly-used public face recongition training dataset always suffer from tlhe long-tailed distribution,di-rectly training models with those dataset may not guarantee good performance.Based on the theory of generative adversarial networks,the author propose a novel data augmentation method by sampling images from long-tailed part,modifying the attribute labels and generating images with modified attributes whose identities remain the same in order to ease the long-tailed distribution.By carrying out experiments for different data augmentation schemes,the data augmentation method in this article reveals significant superiority in face recog-nition performance,proving the effectiveness of the method.