Research On Subtle Facial Expression Recognition Algorithms Based On Deep Learning

The understanding of human emotions and mental states is an important procedure for the realization of artificial intelligence,while the correct recognition of facial expression is the key to judge human emotions and affections.Therefore,facial expression recognition is a key research topic in human computer interaction,and is also an important problem in the field of computer vision.Traditional expression recognition usually deals with facial expression images taken in laboratory conditions.Due to the efforts of academic community,algorithms for this problem have made great progress.Subtle facial expression refers to in-the-wild real expressions with small amplitudes,which can be regarded as hard examples of expression recognition.Traditional algorithms usually fail to explore the smaller discriminative information and the correlation of different local muscle movements.To deal with the above mentioned problem,this paper exploited visual attneion mechanism to model the small muscle movements and their correlation,and tries to extract more discriminative features.The main contents and contributions are as follows:1)A hierarchical attention based subtle facial expression recognition algorithm is designed.The low level,middle level,and high level features of deep neural network contain information of edges,shapes,and semantics,respectively.Therefore,this paper exploits attention module at different layers of neural network to enhance the salient information and suppress noises in different level of features.By this way,the subsequent layers of network are easier to focus on salient information.The algorithm is evaluated in subtle facial expression dataset LSEMSW and the accuracy is 36.73%,which is competitive.The experimental results demonstrate that the proposed hierarchical attention based subtle facial expression recognition algorithm is effective.2)A local-correlation attention based subtle facial expression recognition algorithm is proposed.Firstly,due to the small variations in subtle facial expression,it is difficult to extract discriminative features.Therefore,a group of attention masks are utilized to extract small local features in different local areas respectively,and a regularization is proposed to ensure different attention maps focus on different local facial parts.And then,considering that those small variations are closely correlated,and they together form an expression,therefore the self-attention module is utilized to extract global correlation feature.Finally,an efficient feature fusion method is proposed to reduce feature dimensions and fuse the local features and global correlation feature together for final classification.The algorithm is evaluated in two facial expression datasets.The recognition rate of the proposed algorithm is 38.61% in subtle expression dataset LSEMSW,which is 1.94% higher than former algorithms.In traditional facial expression dataset CK+,the recognition rate is 96.9%,which is 0.5% higher than previous methods and achieves the best accuracy on this dataset.Experimental results demonstrate that the proposed algorithm can better extract weak features and correlation of different local regions,and enhance the recognition rate on subtle facial expression recognition.What's more,the algorithm is able to achieve the best performance in traditional expression dataset,which demonstrates its good generalization power.