Facial Attributes Analysis Based On Deep Learning

Facial attribute analysis is a hot topic in the field of computer vision.Intelligent and accurate facial attribute analysis has wide and deep practical significance.The existing facial attribute analysis algorithms are mainly divided into shallow model based methods and deep learning based methods.The advantage of the shallow model based method is that the model is small,but its disadvantage is that the accuracy and robustness of the estimation are poor,and the ability of feature extraction is limited.In recent years,with the rapid development of deep learning methods based on large-scale dataset training,in the research of facial attribute analysis,the deep neural network has become the current mainstream method because of its powerful ability of facial feature extraction and excel-lent performance.However,in the practical application of the algorithm,the influence of many interference factors(face images with posture,illumination,occlusion and disloca-tion changes)on the whole model is huge,so the design of efficient facial attribute analysis is still of great scientific research and practical significance.To solve these problems,this paper will focus on four important attributes of facial attributes:gender recognition,age estimation,head pose estimation and expression recognition,and propose three algorithms of facial attribute analysis based on deep learning,including:(1)This paper proposes a gender recognition algorithm based on the combination of self-learning and deep learning.The algorithm is based on a convolutional neural net-work.On this basis,the self-paced learning strategy is added,and a gradually learning training strategy is adopted.This model emphasizes more on reliable samples and trains from simple samples to complex samples.With the gradual learning of the model and the addition of complex samples,it avoids The model of deep learning falls into a bad local minimum.The accuracy of the algorithm on IMDB and UTK database is 1.55%and 1.43%higher than that of the depth model without self-paced learning respectively.(2)This paper proposes an algorithm of self-paced deep regression forests consid-ering the uncertainty of samples,which is used for age estimation and head posture esti-mation.First of all,with the help of self-paced learning,the algorithm distinguishes the image of confusion and noise from the conventional image and reduces the interference brought by them.At the same time,by combining the uncertainty of the sample,it reduces the biased solution in the deep regression forests caused by the imbalance of the sample label.The average absolute error of the algorithm in Morph Ⅱ and FG-NET age estima-tion database is 0.26 and 0.29 years lower than that in deep regression forests algorithm respectively,and the average absolute error in BIWI head attitude estimation database is 0.26 degrees lower than that in deep regression forests algorithm.(3)A deep learning algorithm based on visual saliency is proposed for facial expres-sion recognition.This algorithm explains facial emotion expression accurately and effi-ciently through visual saliency prediction and explores the effect of regional information of visual saliency image on the deep model.The accuracy of facial expression classifi-cation and confusion matrix on FER2013 and CK+database verify that the information of visual saliency region can almost reach the high accuracy prediction level of complete information of original image,which shows the effectiveness of visual saliency region.