Research On Facial Attributes Recognition Based On Deep Learning

In recent years,due to the increasing network bandwidth and storage capacity and the rapid spread of various graphics acquisition equipment,the internet has emerged hundreds of millions of facial photos.The rapid increasement in the number of face images provides a great deal of training data for recognizing human identities and attributes(such as age,gender,and expression).How to make full use of these photos for effective facial attributes recognition has become an important hot research topic.The traditional local feature,such as Gabor filter feature,Local Binary Pattern(LBP)and Biologically Inspired Feature(BIF),have some robustness to illumination,posture and occlusion,but their abilities to describe and understand semantics are limited,affecting its performance in actual application of recognizing.In recent years,with the GPU computing performance by leaps and bounds,the computer's computing power greatly enhanced,making the use of large-scale image data training complex deep network becomes feasible.In 2012,Alex Net made the first in Image Net challenge,and since then deep convolution neural networks have made significant progress in multiple areas of image recognition.In the face-related recognition task,the deep convolution neural network has become the mainstream method because of its superior performance.In this paper,we use the convolution neural network architecture in deep learning to study the recognition method of face attribute under uncontrolled condition and develop face recognition system.Based on the multi-task learning,this paper designs or improves the structure of CNN with deep convolution neural network to improve the accuracy of facial attributes recognition.Three kinds of facial attributes recognition methods based on depth learning are proposed,including:(a)An age estimation method combined CNN with latent factor analysis,which embeds the latent factor analysis into the CNN framework.The new sub-decomposing layer can decompose the feature of the convolutional layer into age factor and identity factor.The age factor after the decomposition is used to estimate the age.The contribution of this method is as follows: 1.A new layer: latent factor decomposition layer is proposed,which conjunctes CNN with latent factor analysis algorithms.2.An orderly regression function based on age is proposed,which is optimized with the ordered regression function based on age.3.The method achieves better performance on the MORPH album I,MORPH album II,and FG-NET databases.(b)A gender and facial expression recognition method based on deep learning,which uses multiple patch training to derive multiple models and then fuse multi-model.The method used this method on CVPR Workshop 2016 to achieve the first place on the task of Gender and Smile Recognition and Complement Classification.We also propose a multi-task learning method which regresses the face-to-face key point coordinates and classifies facial expression simultanously.(c)A facial attributes recognition method using a spatial transformer network.The method first uses the location network Lo-Net to learn the parameters in the face-aligned affine transformation,and then uses the new spatial transform layer aligning the original facial photo and finally sents the aligned image into the classification Network Cl-Net for classification.The method increased the average by 4% and 2% on the Celeb A and LFWA databases,respectively.Facial attributes recognition methods based on deep learning proposed in this paper improves the performance on the database of age,gender,and expression and common attributes.Furtherly,the facial attributes recognition methods lay a solid foundation for the facial attributes recognition application system.