Research And Application Of Ensemble Machine Learning Algorithms In Facial Attribute Prediction

Face attribute prediction plays an important role in practical applications such as en-tertainment,security,and social media.In the existing research,the classical machine learning algorithm with better classification and regression performance is used to pre-dict the single attribute of the face.Because the extracted features are shallow and can not effectively mine deeper features,the accuracy of face attribute prediction is difficult to meet the needs of real applications.With the rapid development of computer technol-ogy,deep learning algorithms can efficiently acquire shallow and deep features related to attributes,and can achieve better performance than classical learning algorithms.How-ever,due to the deep layer of deep learning algorithms,the training process is complicated.Moreover,existing public face datasets are difficult to meet such needs.How to design shallow neural networks,extracting effective attribute features,and algorithm classifiers or regressors making full use of these features for final prediction,these are a big chal-lenge in the field of face attributes.Therefore,this paper attempts to explore and combine the unique advantages of deep learning algorithms and classical learning algorithms to design new algorithms to achieve better prediction results.Most attribute predictions of faces are a binary classification problem,such as gender prediction,ethnic prediction,wrinkle prediction,and beard prediction,and the prediction process is relatively easy.Age attribute prediction is a regression problem,which is sus-ceptible to other attributes of the face and external factors.Therefore,age prediction is the most challenging research of all attribute predictions of human faces,and it is also the focus of this paper.In addition,most of the current algorithms only predict a single at-tribute of a face,and cannot predict multiple attributes at the same time.However,there is a strong correlation between face attributes,and excavating the relationship between attributes can enhance a single attribute,which improves the predictive performance of each attribute.Therefore,this paper attempts to explore the correlation between multiple attributes,which is used to enhance the original feature information of each associated attribute,thereby improving the accuracy of multi-attribute prediction.The main work and innovations of this paper are as follows:Firstly,in order to obtain better age prediction performance on face datasets with-out manual processing,an integrated CNN-ELM recognition algorithm is proposed.The Convolutional Neural Network(CNN)adequately extracts valuable attribute information from the picture and uses the Extreme Learning Machine(ELM)for final attribute pre-diction.CNN-ELM is learnt in an end-to-end manner,which makes full use of the good feature extraction characteristics of CNN and the fast and efficient processing capabil-ity of ELM,and achieving better performance on the Adience dataset without manual processing.Secondly,since the CNN-ELM algorithm does not consider the influence of gender and ethnicity on age prediction,an ensemble CNN2ELM attribute enhancement algorithm for predicting age is proposed.It consists of three parts:age attribute feature enhance-ment,age range divided with ELM classifier,and age prediction using ELM regression.Firstly,the extracted gender and race characteristics are integrated into age characteristics.Then,ELM divides them into a certain age range.Finally,we use the ELM regression to predict the final age.Then,CNN-ELM and CNN2ELM algorithms are based on static images for age es-timation,and age is the embodiment of face aging.What's more,the face aging law con-forms to a flow pattern distribution,and if one algorithm can learn the flow distribution well,it can accurately predict the age of the face.Therefore,this paper proposes a dy-namic age estimation method based on edge coordination.Aiming at the mainfold pattern distribution characteristics of aging or younger faces,an age prediction model AR-Net is proposed.AR-Net first uses a generation antagonistic neural network(GAN)to learn the characteristics of the aging or younger flow pattern distribution of facial datasets.The trained GAN is used to generate the aging and younger features of each age group,and these features are used to train the corresponding ELM regression.Then,AR-Net is de-ployed on each edge server to estimate the age of pedestrians.At the same time,the training dataset and prediction model are updated continuously by using the face images collected by edge sensors to ensure that AR-Net has good generalization performance.Finally,the performance of AR-Net is validated by Morph-?,CACD,and face dataset acquired online.Finally,CNN-ELM,CNN2ELM,and AR-Net are used to estimate the face age,but there is a strong relationship between face attributes.This relationship not only enhances the age attribute information,but also other face attribute information associated with the age attribute can also be enhanced.Therefore,this paper proposes a multi-task correlation network algorithm for predicting face attributes.To fully excavate and utilize the corre-lations among attributes,a multi-attribute tensor correlation neural network(MTCN)for face attribute prediction is proposed.Firstly,the network shares shallow image features and distinguishes high-level features of each attribute.At the same time,in the process of high-level feature extraction,each attribute subnetwork extracts the feature information closely related to it from other networks to enhance the original feature information of at-tributes.Then,a new tensor correlation analysis algorithm(NTCCA)is used to project the high-level features of the network into the closely related two-dimensional space.Finally,MTCN predicts the transformed features.Compared with the best methods at present,MTCN achieves the best performance on CelebA and LFWA test datasets.