Research And Implementation Of Facial Expression Recognition Based On Deep Neural Networks

Face expression recognition has been a hot topic in the field of computer vision and machine learning.Because of the subtle differences between human facial expressions,it is a challenging task for computers to accurately recognize facial expressions.In addition,many public face facial recognition algorithms have poor performance for face images acquired under unconstrained conditions,and there is a large gap between the algorithms and the actual application.Traditional algorithms of facial expression recognition generally only use the spatial structure of the image information.However,the expression changes is a dynamic process,with a very clear movement change characteristics,that is,timing information.The information is generally obtained by optical flow algorithm with Multi-frame images,but this algorithm is too complex and the computational resources required is too large.For the above problems,the MDSTFN method based on deep spatialtemporal information fusion is proposed in this paper to improve the accuracy and robustness of facial expression recognition.The main research contents are as follows:Firstly,the common facial expression feature extraction algorithms and facial expression classification methods are investigated,and the classical facial expression recognition systems are also reproduced.The recognition accuracy of each recognition system is compared,and according to the results,the advantages and disadvantages of each system are analyzed in the paper.Secondly,a feature extraction method based on facial expression recognition is proposed.Due to the great success of convolutional neural networks in the field of image recognition in recent years,it is introduced to the field of facial expression recognition in this paper and achieves good performance.Moreover,not only the spatial structure information of expression image but also the temporal information of changing expression is used for expression recognition in this paper.Therefore,the expression recognition system can simultaneously fuse the topological information and timing information to improve the recognition accuracy.Then,a method of using average face instead of neutral face to extract timing information is proposed.The above algorithms are all proposed with the condition that the neutral facial expression images in restricted environment is available.But there are many extreme conditions for facial images in natural scenes.In order to extend the feasibility of the algorithm in this paper,the neutral expression of a specific individual is replaced by the average facial expression based on a large number of faces,which is adopted to expression recognition when there is no neutral expression.Furthermore,compared with the way that extracting features firstly and then fusing,the end-to-end approach in facial expression recognition can reduce the complexity of the system and improve the accuracy at the same time.Lastly,the MDSTFN algorithm proposed in this paper is test on several well-known facial expression datasets,compared with the existing facial expression recognition algorithms.The results show that the MDSTFN algorithm is effective and obviously better than other methods in terms of accuracy.