Research Of Facial Landmark Location And Face Recognition Based On Deep Learning

As an important biological feature recognition technique,Face Recognition is regarded as an emerging tool in criminal investigation,enterprise management,financial treads and intelligent transport system.However,the inadequate location of facial features,lower recognized rate in some scenes still impede the wide application of face recognition.Meanwhile,successive breakthroughs of deep learning technology has attracted increasing number of researchers.Recently,some of them have combined deep learning with face recognition and achieved remarkable results.Based on previous research,we analyze current works combining deep learning with face recognition,and indicate that the multi-task learning based deep convolutional network and the multimodal representation based deep convolutional network would significantly improve the performance of facial landmark location and face recognition respectively.The multi-task learning based deep convolutional network considers facial characteristic point location as a main task,and head gesture detection as an auxiliary task.The robustness of facial landmark location and head gesture are acquired through learning in Deep Convolutional Neural Networks(DCNN).The experiments results show that this method performs better than general DCNN and Cascaded CNN.The multimodal representation based deep convolutional network has two differences when compare with general deep convolutional network.One is using multi deep convolutional network to extract global,local and rendered frontal facial features,including some invariable feature such like gestures and covers.The other is the extracted features are reduced dimensions through Stacked Auto-encoders(SAE)to learn more non-linear feature transformations,not through the traditional Principal Component Analysis method.