Application And Research On Deep-learning Model Compression Method In Expression Recognition

Facial expression is an important way for human to express emotion and conve y intention.The efficient recognition of facial expression has a wide application value in the fields of medical care,fatigue monitoring and smart home.With the improvement of hardware performance and data volume,convolutional neural network is widely used in expression recognition tasks.It relies on a large number of parameters to improve the adaptability and accuracy of the model.However,most of these models are suitable for centralized processing of high-performance servers,which brings high data transmission cost and privacy disclosure risk,so it is difficult to deploy directly in edge devices with limited resources.In order to solve the above problems,this thesis realizes the real-time expression recognition on the edge side by three mainstream deep-learning model compression methods: lightweight structure design,knowledge distillation and channel pruning.The specif ic work and research results are as follows:Firstly,combined with the depth separable convolution and local channel attention mechanism,a lightweight efficient attention network(EAN)is designed,which improves the fitting ability of the model on the premise of strictly controlling the amount of parameters and calculation of the model.Secondly,in the training process of EAN,aiming at the problem that the original one-hot label of expression datasets contain little information and can not reflect the similarity relationship between expression categories,a method based on iterative transfer learning between teacher and student networks is proposed by the idea of knowledge distillation.The EAN is selected as teacher and student networks to construct th e expression recognition framework,and the soft labels marked by the teacher network are optimized through iterative training to assist the training of student network.To strengthen the information transmission in the iterative process,the weight transf er is also introduced between each round of iterative training.Finally,in order to further remove redundant feature information to obtain a more concise model,an improved channel pruning method is proposed for EAN after iterative transfer learning betwe en teacher and student networks.This method is based on convolution kernel L2 norm and batch normalization factor γ design pruning metric and ratio,and adopt local channel retraining strategy to avoid the feature loss of iterative transfer learning process between teachers and students.Experiments show that the model ensures good real-time performance and recognition accuracy,and better realizes the balance between complexity and accuracy of expression recognition model.