Research On Facial Expression Recognition Based On Features In Multiple Regions

Facial expression recognition is an important way of human emotion expression.In-depth study of human facial expressions is of great significance for understanding human inner emotions and achieving human-computer interaction.Today,facial expression recognition technology has achieved success in different fields,but facial expression recognition is still a challenging task.In the real world,each individual has different expression modes for expressions,such as "smile" and "laugh".At the same time,there are "occlusion" problems in facial images.These factors will seriously affect the improvement of facial expression recognition rate.In addition,there are more abundant features in the local areas of the face,which need to be explored and studied.Therefore,for the problem of facial expression recognition mentioned above,in this thesis,we have done the corresponding research work,the main content is as follows:(1)This thesis introduces the research background and significance of facial expression recognition,and briefly describes the current status of facial expression recognition at home and abroad.This thesis analyzes and summarizes various facial expression recognition methods.Then the image processing method for facial expression recognition is introduced.Finally,we focus on various facial expression recognition methods based on deep learning,and classify these methods.(2)Therefore,a new LFA(Locality-Feature-Aggregation)loss function is proposed,which can reduce the difference between images of the same kind and expand the difference between images of different kinds during the training of deep neural network,so as to weaken the influence of expression polymorphism on features extracted by deep learning.,At the same time,local areas with rich expressions can better express facial expression features.Therefore,a deep learning network framework incorporating the LFA loss function is proposed.The framework extracts local features of facial images for facial expression recognition.Compared with other methods,the method proposed in this thesis has achieved 87.61% and 96.92% recognition accuracy on the real world RAF data set and the CK+ data set in the laboratory,which fully demonstrates the effectiveness of this method.(3)In this thesis,a deep learning network architecture MRF-CNNs(multi-region feature fusion CNNs)with attention mechanism is proposed.The framework makes full use of different regions of the whole face to extract local features with rich emotions.Meanwhile,the method effectively complements the whole and local features.Then,by assigning different weights to the different features extracted by the MRF-CNNs subnet,we adaptively adjust the importance of different parts of the face,and focus on local features that are not "occluded",so as to solve the problem of " occluded " in facial expression recognition.Compared with other latest methods,the MRF-CNNs framework in this thesis has achieved good recognition effect on RAF data sets and SFWE data sets.