Research On Facial Expression Recognition Method In Natural Scenes Based On Deep Learning

As technology keeps advancing,people’s demand for device intelligence is getting higher and higher.Emotion is one of the highest level of human intelligence,and human emotions can be recognized intuitively from facial expressions,so facial expression recognition naturally becomes a hot spot for research.Facial expression recognition technology has a wide application prospect in the fields of assisted teaching,assisted driving,medical and escort robots,etc.However,the performance of facial expression recognition in natural scenes cannot meet the realistic demand,which is the urgent problem to be solved in facial expression recognitionAt present,face expression recognition in natural scenes mainly uses a single deep learning algorithm,but due to the problems such as complex background interference factors in natural scenes,the recognition accuracy of face expression recognition in such scenes is low and the robustness is poor.To address this problem,considering the advantages of different algorithms in different scenes,this thesis proposes a Q-learningassisted network based on deep Q-learning for static face expression recognition in natural scenes,using Res Net-18,Res Net-50,VGG-16 and VGG-19 as the backbone networks,and selecting suitable expression models for different expression types based on the information of the backbone network recognition results.We select the appropriate deep learning models for expression recognition based on the result information of the backbone network.Through the experimental validation on RAF-DB dataset and FERPlus dataset,it is proved that the proposed algorithm model has higher recognition accuracy and robustness.Based on the static face expression recognition in natural scenes,this thesis further investigates the dynamic expression recognition in natural scenes.To address the problem that dynamic expression recognition in natural scenes is less studied and the existing studies seldom utilize the temporal information of expression recognition effectively,memory Q-learning assisted network is proposed for dynamic face expression recognition in natural scenes.In this model,two modules,long-time memory unit and short-time memory unit,are proposed to be added for fusing expression information of different temporal sequences to improve the utilization of temporal information.The short-time memory unit in this algorithm focuses on the expressions in the current video frames,effectively increasing the proportion of peak expressions while reducing the interference brought by poor quality video frames,while the long-time memory unit focuses on the recognition results on the whole video sequence and fuses the recognition results in all videos,solving the problem of little use of temporal information in the current dynamic expression recognition research.The effectiveness of the algorithm is verified through experiments on the CK+ dataset and AFEW dataset,and comparative analysis with other methods.Finally,a face expression recognition system is designed and developed to validate the proposed Q-learning-assisted network algorithm and memory Q-learning-assisted network algorithm.The system has the functions of static expression recognition,dynamic expression recognition and real-time expression recognition,and provides intuitive expression recognition results,as well as the functions of registration and data query,which can be used for expression recognition in most scenarios with high recognition accuracy.