Multi-task Face Attribute Recognition Based On Deep Learning

In recent years,people's identity authentication methods in life have gradually evolved from traditional means to biometric level with the continuous improvement of artificial intelligence technology and deep learning technology,which makes them utilized widely in finance,entertainment,transportation,community interaction and many other domains.Especially in the research of face-related subjects,face attribute recognition has received special attention due to its own research value and practical value.From the recent research work at home and abroad shows that feature sharing among multiple face attributes with positive correlation will help to improve the accuracy of each recognition task and achieving face attribute recognition tasks with the convolutional neural network has higher reliability and accuracy than traditional feature extraction methods.Consequently,this thesis proposes a hard parameter enhanced channel information sharing method based on deep learning to study multi-task face attribute recognition.The main work of this paper is as follows:Firstly,we conducted a literature survey on the multi-task face attribute recognition work and related basic theory based on convolutional neural network at present and learned about the research status at home and abroad as well as the shortcomings of subject.The hard parameter enhanced channel information sharing based on deep learning was proposed to study the multi-task face attribute recognition.Secondly,the research on multi-task face attribute recognition based on single-task network is carried out.Three single-task network of Alex Net,VGG Net and Res Net is used to construct the feature sharing structure by setting the improved method of multi-tasking output layer to realize multi-task face attribute recognition based on feature sharing respectively.We use recognition accuracy,model size and running speed three performance indicators to evaluate models.The experimental results show that the improved method has certain feasibility,and the Alex Net network performs well(the average recognition accuracy is 92%,85%,85%,respectively),but there is still room for improvement because its performance indicators is not ideal.Thirdly,a convolutional neural network method based on enhanced channel information sharing is proposed to realize multi-task face recognition.The core technology for enhancing channel information sharing is group convolution and sharing algorithms between channels.In this paper,we studied two different network structures and utilize 7 types of face attributes selected on the CelebA dataset to carry out recognition and the average recognition accuracy of them is over 94%.Significant improvements is made in structure and speed by one of models with 0.98M model size and 6ms recognition speed.Finally,the data statistics and data enhancement methods are researched and the PC-side application software is developed.By constructing the GAGR-A dataset,a multi-task face attribute recognition APP based on the convolutional neural network method of enhanced channel information sharing is developed on the Android platform to realize 8 types of task face attribute recognition at the same time.After the data set and laboratory members actually tested,the accuracy rate was 92.7%,the model size was 1.04M and the speed was about 6.7ms.In summary,the multi-task face attribute recognition based on the deep parameter-hardened channel information sharing method can not only improve the recognition accuracy,but also greatly accelerate the recognition speed because of the weight reduction of the model.The foundation of practical application was laid by research.