Research On Baby Face Attributes And Action Recognition Based On Computer Vision

In recent years,face attributes and action recognition have become the focus of research in computer vision.They are not only applied in the fields of intelligent monitoring and scientific parenting,but also in human-computer interaction,business recommendation and social network.However,most of the existing research focuses on the analysis of adults' faces and actions.As we all know,0-2 years old is the golden period for the growth of babies.The research on the recognition of face attributes and action recognition of 0-2 years old babies has important theoretical significance and application value for assisting the monitoring of the growth of babies.With the help of computer vision technology,this paper will study the face attributes recognition and action recognition methods of 0-2 year old babies.The specific research contents are as follows:Firstly,in view of the deficiency of the existing face attribute data set,in this chapter,a new face image dataset for babies aged 0 to 2 years old,named Baby Face,is established.Afterwards,for the problem of imbalance of baby facial expression data,the Sad and Happy facial expression images were amplified based on the Cycle GAN method.Secondly,according to the different characteristics of baby expression and adult expression,a new method of baby expression recognition based on feature guidance and optimized class spacing is proposed.First use the pre-trained VGG16 network to extract features on the Baby Face expression data set,then design a suitable attention mechanismSqueeze-and-Excitation(SE)module to portray important information for expression recognition,and then use a Second-order Covariance Pooling method better characterizes the area distortion of key points in the image.Finally,a novel loss function called Distance Loss is designed.Afterwards,the effect of different methods on the recognition of the baby's age is verified,which verifies that the method based on feature-guided optimized class spacing(VFESO-DLSE)has the best accuracy on the recognition of the baby ‘s facial expression.Thirdly,aiming at the problems of baby age recognition and gender recognition,an baby-oriented gender and age estimation algorithm is designed.First we design the age estimation network structure,adopt the 2-Stream structure,and use multiple stages of coarse-to-fine multi-level classification and add attention to the network structure design.For the gender estimation algorithm,inspired by the age estimation method,the gender estimation also uses the 2-Stream structure.In addition to the conventional data enhancement methods,Random Erasing and Mixup data enhancement are used.Finally,the analysis verifies that our proposed methods have greatly improved the identification of baby's age and gender.Fourthly,to explore the small sample problem caused by the difficulty of obtaining the baby action recognition dataset,through the idea of transfer learning,transfer the knowledge learned from the initializing behavior recognition network in Kinetics data set to the baby action recognition network.When training the baby behavior recognition network,3D Res Net series and 3D Res Ne Xt are used to train the parameters on the Kinetics dataset to initialize the behavior recognition network.The final experiment verified the effectiveness of 3D Res Ne Xt-101 for baby action recognition.Based on the above research content,this paper designs and implements a baby face attribute recognition and action recognition system.The system encapsulates the expression recognition,age recognition,gender recognition and action recognition algorithms proposed in this article into an easy-to-operate executable software,which has been verified by testing each function.