Research On Facial Expression Recognition Based On Feature Fusion

The development in the field of human-computer interaction has been greatly enhanced with the rapid development of computers,and the accurate recognition of expressions is becoming an extremely important interaction technology in the era of human-computer intelligence.After several years of development,the ideas and methods of deep learning have been applied to face feature extraction at a deeper level with very satisfactory results.This paper fully understands the ideas of deep learning,and then aims to improve the shortcomings and deficiencies of the currently existing algorithms related to the field of face expression recognition from the perspective of feature fusion.The main research content and findings of this paper are summarised as follows.A face expression recognition model MVRes Net-FER with multi-view feature fusion under deep residual convolution is proposed to address the problems of inaccurate and computationally intensive face expression recognition under multiple views in real life.The model improves the residual block in Res Net and uses a deep separable network instead of a conventional convolutional network.A CBAM module is added to enhance the extraction of effective features and the supplementation of shallow feature information under multiple views.The original Re Lu activation function is also used instead of the original Re Lu to avoid partial node deactivation.Finally,the global average pooling layer is used instead of the fully connected layer for the dimensionality reduction operation,and the generated feature vectors are fed into the Softmax model for analysis and classification.This method not only enhances the extraction of effective features and the supplementation of shallow feature information under multiple perspectives,but also avoids the deactivation of some nodes when the gradient is large and improves the classification accuracy.To address the problems of small samples of real-life tagged data and less accurate recognition,a semi-supervised expression recognition model MSFVGG-CW with multi-feature fusion embedded Chinese Whispers(CW)clustering is proposed.this model improves the VGG-16 network by extracting global features and local features for the whole face,eyes and mouth respectively,while combining compression and reward and punishment network modules to reduce the loss caused by different occupancy ratios during network training,then the extracted facial feature information is fused,and finally a dimensionality reduction operation is implemented and the fused features are fed into CW clustering for clustering analysis.The model takes into account both overall and local feature extraction of the face,which can effectively reduce the loss generated by network training and improve the accuracy of expression recognition.An SVM expression recognition algorithm incorporating parallel networks is proposed to address the problems of traditional convolutional networks with simpler feature extraction functions and weaker functions for model generalisation.The algorithm improves CNN and VGG-16,and then takes feature extraction operation on the input face expression image,improves the extraction ability of shallow features and deep features,and fuses the features of the extracted parallel networks,and finally uses the support vector machine method to obtain the category labels of the tested image by prediction.This method improves the generalization ability of the model,and the non-linear operation complements the feature information and facilitates the acquisition of fused image features that are easier to classify.