A Study Of Facial Landmark Detection And Facial Expression Recognition Based On Deep Learning

Facial expression recognition algorithm has important theoretical value and practical significance in the field of computer vision.In recent years,with the continuous development of artificial intelligence,facial expression recognition has been widely used.With the development of deep learning,convolutional neural network has gradually become one of the main choices for most tasks in computer vision.The accuracy of facial expression recognition is usually affected by the facial occlusion and facial angle.In response to these problems,this thesis first employ the facial landmark detection algorithm to align and divide the face,and then introduces the attention mechanism to improve the accuracy.The main contributions are summarized as follows:(1)The existing algorithms are investigated.The current research status of deep learning convolutional neural network algorithms is introduced.The problems and advantages of applying deep learning to these tasks are discussed.(2)In the study of facial landmark detection algorithm,a path-aggregation network structure is introduced based on the network built by residual module.It can promote the flow of information in each layer network.Then,in order to improve the ability,a facial landmark detection method based on heatmap is employed.In the 300 W evaluation dataset,the error was reduced by 0.85.The accurate facial landmarks are helpful to extract the face expression features that are more robust to partial occlusion and angle.(3)In the study of face recognition algorithm,in order to solve the inaccurate caused by partial occlusion of face,the attention mechanism is introduced.It includes attention mechanism between convolutional neural network channels and between image regions.The interregional attention mechanism contains the self-attention mechanism and the correlationattention mechanism of image regions.Then,a series of methods such as mix up and label smoothing are adopted in to solve the over-fitting problem caused by the insufficient data in the datasets.The accuracy of FER-Plus and RAF-DB datasets was improved by 2.27% and 2.03 %respectively.The experiment results show that the proposed facial landmark detection model based on heatmap can improve the positioning accuracy.The expression recognition algorithm based on the attention mechanism can improve the recognition accuracy on FER-Plus and RAF-DB,especially on the data subset composed of occlusion faces.