Facial Action Unit Detection Based On Deep Learning

Facial expression recognition is a very important module of face research.For example,it can recognize the peoples' emotions to avoid emergencies in the intelligent security scene,and it can identify the suspect's microexpression to assist the case handlers to judge whether they lie or not during the investigation and interrogation.It also can capture the students' facial expressions to judge whether the students have doubts about the content of the lecture.But only six basic expressions are not enough to describe the responsible human emotions.To solve this problem,we choose facial action units,which are more basic units of facial expressions.The combination of these facial action units can be used to describe more facial expressions.Based on deep learning,this paper studies the detection methods of facial action units,and proposes two detection algorithms of facial action units.The main research results of this paper are as follows:(1)Aiming at the problem that some facial action units occupy a small area of the face area,which leads to the difficulty of detection,NLCA Block is designed,which is a non-local channel attention module.The module is embedded into ResNet50 and Efficientnet-B3 network structure by using the portability of Non-local Block.Because it's difficult to train the two networks from scratch,we use migration learning to recover the convolution parameters in the model trained on ImageNet,and then use the dataset of facial action units to finetune and complete the feature extraction of facial action unit.The experiments show that NLCA Block has the advantage of acquiring context semantics in the shallow part of the network,which is helpful to improve the accuracy of small target detection.In addition,the co-occurrence relationship between the different AUS is implicitly studied by using its advantage of capturing long-distance features.The two models designed in this paper are tested on BP4D and DISFA datasets.It is also proved that the models are better than some existing facial action unit detection models and has good generalization ability.(2)To solve the problem of unbalanced distribution of AU samples in facial action unit datasets and difficult-classified samples,focal loss is proposed to solve this problem.The experiments are carried out for multiple values of two parameters for focal loss.In addition,the loss function of AU multi-label co-occurrence is proposed to implicitly learn the relationship between AUs.The algorithm has achieved good accuracy in both CK+ dataset and BP4D dataset.It is proved that the loss function of focal loss and AU multi-label co-occurrence relationship are both effective to improve the experimental accuracy in the detection task of facial action unit.