| Facial Expression Recognition(FER)refers to the process of computer recognition of expressions by extracting and analyzing facial features.It has great social needs and broad application scenarios in many fields,such as emotion calculation,behavior prediction,human-computer interaction,mental health services and so on,and its research has been widely concerned.In the real scene,the problem that the accuracy of facial expression recognition decreases due to partial occlusion of the face by glasses and hands is an urgent research difficulty.Therefore,this thesis mainly studies the robust facial expression recognition under natural occlusion.Feature is the main basis of facial expression recognition,and occlusion brings interference features and makes some discriminative facial features lost.Therefore,optimizing feature extraction is an effective means to improve the robustness and accuracy of occluded expression recognition.In recent years,deep learning method has shown strong learning ability and discrimination ability,and has become a common research method in the field of expression recognition.Therefore,this thesis takes deep learning as a tool and optimizing feature extraction as a breakthrough,and proposes a method based on deep learning network to optimize feature extraction and improve the recognition accuracy of occluded expression.The main research contents and innovations of this thesis are as follows:(1)In order to eliminate the interference of the occlusion area,this thesis proposes an occluded expression recognition method based on attention guidance,which guides spatial attention by an occlusion position information,thus constraining the generation of expression features.Firstly,based on facial landmarks,an occlusion indicator is designed to obtain the location information of occluded areas and key areas without occlusion,and an attention descriptor is designed,which reflects the attention of the network to the specific spatial location of features.Then,the attention loss is constructed on this basis,so that the occlusion indicator guides the attention descriptor.Finally,the overall loss is constructed by further fusing the loss of expression classification,so that the depth features have accurate attention adapted to specific expressions,so as to improve the accuracy of expression recognition.The results of feature visualization and expression recognition on five public occluded facial expression datasets show that the features optimized by attention guidance strategy have accurate attention,and the performance of expression classification is better than that of existing methods.(2)In order to suppress the interference of occluders and enhance the expression ability of discriminative expression features,this thesis proposes a Group-level Feature Optimization(GFO)module from the semantic feature level,which can make the semantic features disturbed by occlusion are well distributed again and then converge into high-quality overall features.The GFO module groups features in the channel dimension,and the feature vectors of each spatial position in the feature group are adaptively weighted by their similarity with the group-level core semantics and their relevance with the global core semantics.Finally,the GFO module supplements the group-level features with global information fused with channel attention.GFO module is embedded in all levels of ResNet50 from shallow to deep,forming the occluded expression recognition network GFO-ResNet50.By making full use of group-level and global information,it realizes self-adaptive mining of core detail features,integration of related semantic features into groups,and iterative optimization of group-level features.At the deepest level,group-level semantic features with good distribution are synthesized into high-quality features for occluded facial expression classification.In this thesis,a large number of experiments are designed on five public data sets of occluded facial expressions.The results of semantic visualization and expression recognition show that GFO module optimizes semantic features,and the performance of expression recognition is better than that of existing methods. |
