Research On Facial Expression Analysis Based On Conditional Generative Adversarial Nets

Facial expression is a direct and noverbal reflection of human emotions.It is also an important way for people to read other's feelings.Facial expression analysis is the basis for understanding human emotions and the prerequisite for man-machine interaction.In the field of computer vision research,facial expression recognition has drawn wide concerned as an important part of analyzing human emotions.Only by recognizing human facial expressions can computers understand human emotions and further communicate with humans.In recent years,with the rapid advancement of deep learning,research on expression recognition has also made breakthrough progress.At present,many expression recognition methods have achieved a high recognition rate for the six basic expressions,but there are still some difficulties in recognizing the complex facial expressions of humans in reality.First,facial expression recognition is susceptible to interference from individual identity information.For example,human facial features,skin color and other information will weaken the express of expression information to a certain extent,and therefore the expression analysis results will be disturbed.Secondly,there exist differences in the intensity of expression.These problems make it a great challenge to analyze the expression of the computer accurately.In response to the above problems,this thesis proposes a method of expression classifi-cation and intensity estimation grounded on conditional generation adversarial networks to improve the accuracy and robustness of expression recognition with varying intensity.First,the collected facial expression pictures are divided into different expression categories and intensity.Then combine the expression pictures with the neutral expression pictures into pairs for training.Next,the expressions are classified and estimated intensity.The main work of this thesis are:(1)Proposed a neutral expression generation method based on conditional generative ad-versarial network(cGAN).This thesis uses conditional generation of adversarial networks to learn the mapping relationship between expression features and neutral expression fea-tures,and to convert facial pictures with expressions into corresponding neutral expression pictures.In the third chapter of the thesis,through the subjective human eye observation and objective parameter analysis,the generated neutral facial expression pictures are evaluated.The results prove that the generated pictures and the authentic pictures have high similarity.(2)Proposed a facial expression analysis algorithm combining facial expression clas-sification and intensity estimation.The algorithm differentiates expressions in the trained generation network to realize the extraction of expression features.Then use classifier to classify the category and intensity of the expression.In order to estimate the intensity of the expression,it is necessary to define the intensity of the expression.The first frame in the expression video sequence is defined as a neutral frame,the 7-9 frames are set as weak,and the last three frames are peak.The algorithm not only categorizes 6 basic expressions,but has the intensity of expressions estimated,which can effectively improve the accuracy of expression classification.(3)The algorithm proposed in this thesis is compared with the existing facial expression analysis algorithm on several public facial expression datasets.The results show that the expression analysis method proposed in this thesis is significantly better than other methods.In terms of recognition accuracy,it is higher than the accuracy rate of existing expression recognition,and greatly improves the recognition accuracy of weak expressions.In terms of intensity estimation,the algorithm not only improves the recognition accuracy of peak and weak expressions,but also promotes the effect of expression recognition.Then on the basis of this research,it analyzes future research difficulties and research directions.