The Study Of Facial Expression Recognition Method Based On Deep Learning And Arousal-Valence Model

With the advent of the era of artificial intelligence,AI seeps the various trades and occupations comprehensively.Human expect computers not only to simulate human intelligence,but also to describe human psychological activities.Facial expression has been a hot area of artificial psychological research.Through computer studying,psychological research has received more and more attention.The study is carried on deep learning and Arousal-Valence emotion model.The main research is focus on Gabor wavelet and convolutional neutral network features fusion based on multi-output regression method,besides another method is based on fusing convolutional neural networks multi-layers features.Firstly,Gabor wavelet has the advantage of ability in describing the local features of image,convolutional neutral network can learn global features of image.Fusing image's Gabor wavelet features and convolutional neural network features,to make their respective advantages complementary to each other.Then using the correlation between arousal and valence to achieve multi-output regression.The experiment results indicate that the proposed method is better than traditional methods.Secondly,because different convolutional layer has different learning effects for image,shallow layer features have more detail information,deep layer feature have more abstract semantic information.In order to acquire better facial features and use the correlation between arousal and valence,a method of based on fusing convolutional neural networks multi-layers features and sparse representation is proposed in this thesis.This method has been fully explored the features convolutional neural networks,using sparse representation algorithm to achieve feature selection in convolutional layers.The sparse representation algorithm calculates the weight between feature map and label,choosing the more relevant features with labels,thus dimension reduction is achieved.Then fusing the treated convolutional layer features,put the new fusion features into support vector regression to train and predict.The contrast experiments show that the proposed method is better than deep learning method and other non-deep learning method.