Research Of Facial Attributes Recognition Based On Convolutional Neural Networks

Facial Attributes Recognition means using computers to identify and analyze a variety of facial attribute information in images,including biological recognition and non-biological recognition.The former includes the determination of general biological factors such as age,gender,race and specific biological factors like hair color and eyebrow shape,besides facial expression for instance smile,anger and etc.Non-biological determination contains the identification of accessories like glasses and hat.Facial Attributes Recognition has raised growing attention since its wide use on Face Verification,Face Identification and Face Retrieval in recent years.The traditional methods for Facial Attributes Recognition generally follow the procedure that extracting facial characteristic artificially firstly then training the classifier.It is well known that the generalization of classifier is subject to the representation of facial characteristic.Facial Attributes Recognition under unrestraint condition is still a challenged research subject because the traditional methods mentioned earlier failed to respond to facial changes influenced by lightings,poses and occlusions resulting from that manual method is only able to extract low-level facial features limited in visual concept.However,the automatic learning of high-level facial features in semantic concept with self-learning ability of Convolutional Neural Networks becomes to be possible based on the development of Big Data and Cloud Computing technology.In conclusion,this thesis mainly researches on the Facial Attributes Recognition based on Convolutional Neural Networks(hereinafter referred to as CNNs).Firstly,this thesis explains the background and significance of CNNs combining the summary of existing research on Facial Attributes Recognition.Secondly,this thesis introduces the traditional method of Facial Attributes Recognition and basic theory of Artificial Neural Networks and CNNs.Thirdly,this thesis proposed a deep CNNs with robustness towards complex background and facial changes for multi-label Facial Attributes Recognition under unrestraint condition,which received two prediction accuracy rate higher than that by traditional Facial Attributes Recognition method on 9.638%and 7.605%separately.Lastly,this thesis proposed two strategies for adjusting the network architecture,which aim at status quo that the training dataset related to age and gender is narrow.The average prediction accuracy generated by two strategies mentioned before all exceeded that by most advanced method of age and gender classification.