Research And Application Of Lightweight Face Expression Recognition Model For Embedded Devices

With the continuous expansion of artificial intelligence application fields and the continuous development of deep learning technology,the way to recognize and analyze human language,expressions,actions and other communication information through deep learning technology has shown a diversified development trend.Expressions are the concentrated reflection of human psychology on the face,which can convey people’s emotions and feelings at that time,and are an important means of communication.In practical applications,embedded devices have the advantages of low power consumption and easy application,and deploying facial expression recognition models on embedded devices can make the application scenarios of facial expression recognition more flexible.At the same time,the offline deployment of models on embedded devices also solves the network communication limitation and user’s privacy and security problems brought by the cloud deployment.The current mainstream facial expression recognition models usually have complex network structures and large number of parameters,which are difficult to deploy and run on embedded devices.Therefore,this thesis investigates lightweight face expression recognition models deployed on embedded devices,and conducts corresponding deployment and application validation with Huawei Atlas 200 DK to assist in the localization implementation of embedded artificial intelligence.The main works are as follows.(1)To address the problem of insufficient and unbalanced data volume in the face expression dataset,this thesis introduces a generative adversarial network to perform data augmentation on the face expression dataset.This thesis increases the volume of face expression data and alleviates the differences in the volume of data for each category of expressions to improve the generalization performance of the face expression recognition model.In this thesis,the Star Gan network is used to enhance the face expression dataset like RAFD and RAF-DB by generating face expression images with similar data distribution as the original dataset,and improve the quality of facial expression generation of Star Gan through cross-dataset migration learning.It is demonstrated that this approach can effectively enhance the performance of the face expression recognition model.(2)To address the problem that the mainstream facial expression recognition model has a large number of parameters and computation,which is not suitable for deployment on embedded devices.This thesis designs a lightweight facial expression recognition model Light FER based on the Mobile Vi T framework,and combines CNN and Transformer methods to design the facial expression recognition model,which enhances the performance of the facial expression recognition model for face.The model is designed to enhance the ability of the facial expression recognition model to extract local features and establish connections between local features in the facial expression images.This thesis introduces channel attention mechanism and random feature discard mechanism to enhance the extraction ability of the model for each channel of the face expression image and force the face expression recognition model to pay attention to the association between different local features.It is experimentally verified that Light FER has low number of parameters,low computational effort and high performance,which is favorable for deployment on embedded devices.(3)The deployment of Light FER based on Atlas 200 DK embedded devices is completed,and the feasibility and practicality of Light FER on embedded devices is verified.The embedded facial expression recognition system in classroom environment is also implemented,which can be used to assist teachers to evaluate the quality of classroom teaching.