Research On Facial Expression Recognition Technology Based On Feature Fusion And Deep Learning

With the rapid development of artificial intelligence,the importance of facial expression recognition is becoming more and more obvious.It has broad application space in safe driving,intelligent human-computer interaction,case detection,game innovation and other fields,and has become a research hotspot.The traditional expression recognition method is complex in feature extraction,which is greatly affected by the environment and has low recognition accuracy;the deep learning expression recognition method has insufficient feature expression ability,low recognition accuracy and large amount of model parameters.In this paper,facial expression recognition method is studied based on feature fusion and deep learning,the main research contents are as follows:1.The quality of the features extracted from the model directly affects the effect of expression recognition.Aiming at the problem that the feature difference between different expression data categories is small and the features extracted from the network are lack of pertinence,this paper proposes a method to enhance the image feature information in the data preprocessing stage.Local binary pattern and discrete cosine transform are used to obtain the local features and global features of the facial expression image respectively,and the two groups of features are weighted and fused to reconstruct the data set,so as to enhance the facial expression related feature information and reduce the interference of redundant information;the enhanced data set is classified and recognized by residual network,so that the feature extracted from the model has stronger expression ability.Compared with the original expression data,the recognition rate on the JAFFE data set increased by 0.45%,and on CK+ decreased by2.02%.The effectiveness and practicability of this method need to be improved,and further research and improvement are needed.2.Aiming at the problems of the existing convolution neural network model algorithm,such as the lack of feature extraction ability,the low recognition accuracy and the large amount of model parameters,a residual network facial expression recognition method integrating attention is proposed.Firstly,convolution block attention module is improved by feature fusion method to enhance its ability to enhance the key features of expression;the improved attention model is integrated into the residual network,and the discriminant expression of expression features is enhanced in the model feature extraction stage,and then the expression features are extracted by convolution layer to improve the performance of model feature extraction.Through the experimental comparison,after integrating the attention into the network,the model can effectively improve the recognition rate with a small increase of model parameters;after the improvement of convolution block attention module,the recognition effect and stability of the model are further improved,and the recognition rate is improved by 2.68% and6.06% respectively on FER2013 and CK+ data sets,which proves the effectiveness of the method.3.Through the further study of attention mechanism,aiming at the problems of large amount of parameters and poor performance of current attention model,a super lightweight dual pooling channel attention model is constructed.The model uses one-dimensional convolution to capture the cross channel interaction information,and uses the method of two channel feature vector addition and fusion to enhance the performance,and optimizes the channel importance weight acquisition method.The experimental results show that the model can improve the recognition effect of CK+ and JAFFE by 4.14% and 6.07% with negligible parameter increase,which verifies the effectiveness of the channel attention model.The channel attention model is combined with spatial attention to build a super lightweight dual attention model,and the performance of the model is stabilized by adding shortcuts.Compared with the improved convolution block attention module,the parameters of the model are reduced by nearly 3M,and the performance of facial expression recognition is stable and slightly improved.