Age Estimation Based On Integrated Convolutional Neural Network

At present,face recognition technology has matured and is widely used in many fields.Compared with face recognition technology,the age estimation technology is relatively immature.In the meantime,the social demand for face age estimation is getting bigger and bigger.A face image is a biological feature.It identifies an individual by identifying the biological feature.However,due to the differences in human individual living habits,working environment,innate genes and other factors,individuals of the same age show great differences in the external characteristics of age.So the research and development of age estimation still faces enormous challenges.The traditional way to extract facial age features is manual.The process is cumbersome and only shallow features.Convolutional neural network is a commonly deep learning algorithm.It directly uses the image as the input of the network.And it automatically extracts the deep and discriminative feature through layer-by-layer operation and backpropagation.In order to avoid the excessive pre-processing and manual selection of features,as well as more fully extract features,this paper combines the image processing method of convolutional neural network to study the age estimation.In daily life,different people show different degrees of aging due to some factors such as living habits,working environment,and genes.Therefore,under current conditions,face age estimation efficiency is low.Convolutional neural networks are one of the best algorithms for processing image problems today.The current classifiers for age estimation are generally carried out separately.Therefore,there is a lack of information exchange between classifiers.The paper is based on the idea that the whole is better than the part.Then,it is studies the face age estimation along the integrated learning context.First of all,it involves collecting and preprocessing digital images of faces.Then,it extracts facial age features and generates a number of different base classifiers through two different convolutional neural networks.Secondly,it uses an integrated algorithm to combine multiple base classifiers so that information between different base classifiers is communicated.Finally,the integrated classifier is used for age estimation.The main research contents are as follows:1.Image PreprocessingIn order to increase the generalization ability of the dataset to the network,we have expanded the dataset tenfold.Amplification methods are multi-angle cropping and mirroring.In addition,the face image has different illumination intensity and various head postures,which directly affect the generalization ability of the network.To reduce the impact caused by environmental changes,this topic preprocesses the image of the data set before network training.The preprocessing method is histogram equalization.2.Research on Age Feature Extraction of Face ImageIt extract facial age features by GONET with dual-channel and multi-convolution kernel and FYNET with 1×1 convolution,channel shuffling and sparse connection.Then it obtains the base classifier trained on the FGNET dataset and the CACD2000 dataset.The GONET network is designed to improve the accuracy of the network and reduce parameters to prevent overfitting.GONET design is based on the AlexNet and does not change the input and output of the network convolutional layer.The depth and nonlinearity of GONET are increased by using the equivalence between 1×3 convolution,3×1 convolution and 3×3 convolution.GONET enhances the ability to express image features.The experimental results show that the GONET network recognition effect is significant.GONET parameters can be greatly reduced while improving the recognition accuracy.The FYNET network aims to maintain the accuracy and miniaturization of the network.The feature information between different channels is exchanged by channel shuffling.The information between the layers and layers is better fit the age characteristics to improve the network's ability.At the same time,1 × 1 convolution and sparse connections can greatly reduce the parameters of the network.The experimental results show that the FYNET network can realize the miniaturization of the network model in the case of maintaining the accuracy of the age estimate.3.Research on age estimation integrated algorithmIn order to enhance the information exchange between single classifiers,three integration strategies are used between the base classifiers,the maximum probability voting method,the global probability weighted average method and the maximum probability weighted average method.The base classifier is generated by GONET and FYNET.The maximum probability voting method is to obtain the integrated output by counting the results of all base classifiers.The weighted average method of global probability and the weighted average method of maximum probability are obtained by taking the output of the base classifier as the input of the algorithm.Then it gives each base classifier a certain weight.Finally,the integrated output is obtained according to the weighted average method.The weight update method is the output probability of all categories based on the base classifier and the maximum probability of a single category.Experiments were carried out on the age datasets FGNET and CACD2000.The experimental results show that the integrated classifier can combine the advantages of a single base classifier and obtain a better age estimation effect than a single base classifier.