Research On Facial Expression Recognition Based On Generative Adversarial Networks

Facial expression recognition algorithm has been widely concerned by researchers.With the advantages of high efficiency and easy training,deep learning technology has been commonly used in the field of image processing,especially in the field of facerelated.The emergence of convolutional neural network greatly improves the recognition rate of face recognition and facial expression recognition.However,deep learning technology relies on a large number of training data.There are some problems in the open database of facial expression recognition,such as small sample size or unbalanced classification.On this basis,the training model easily emerges over-fitting,which limits the development of facial expression recognition.Data enhancement is an effective way to solve the problem for small sample and imbalanced database.As an alternative evolution of neural networks,Generative Adversarial Networks(GANs)has shown great practical value in image generation.This thesis focuses on data enhancement and classification of facial expression images.The specific work is described as follows:(1)Traditional data enhancement methods such as rotation transformation and noise addition cannot solve the problem of high similarity.So in this thesis,we improve CycleConsistent Generative Adversarial Networks(CycleGAN)and constructs a Constraint Cycle-Consistent Generative Adversarial Networks(CCycleGAN)by introducing classconstraints,which implements one-to-many data category mapping and reduces the costs of model training.At the same time,in order to improve the efficiency of facial expression recognition,we replaced the discriminator of CycleGAN by an auxiliary facial expression classifier.This improvement not only discriminates the authenticity of the input image,but also classifies the expression.The experimental results on the CK+ and FER2013 datasets show that the proposed method can effectively solve the problem of over-fitting and sample imbalance,also improve the quality of generated images,thereby improve the recognition rate of facial expressions.(2)When we employ CCycleGAN to carry out data enhancement,a pair of generator and discriminator should be trained,which results in a large amount of parameters.This thesis proposed a facial emotion recognition method based on Contextual GANs to alleviate the complexity of model.CCycleGAN is improved by introducing the contextual loss function,using a generator and a discriminator to build a contextual generation adversarial networks,realizing the mapping from neutral expression to six basic expressions,and enhancing the facial expression databases.Firstly,the image quality obtained by different methods is compared and analyzed.Then,the influence of different loss functions on network training is analyzed by visualizing the output characteristic graph of each network layer in the process of network training.Finally,in order to verify the validity of the image generation method proposed in this thesis,the database is expanded by using the generated image and facial expression recognition is carried out.The experimental results on FER2013,CK+ and KDEF databases show that the contextual loss function can effectively improve the network's ability to extract facial features,so as to obtain higher quality generated images.Data enhancement based on this method can improve the facial expression recognition rate on The KDEF data set.