Research On Facial Expression Recognition Based On Convolutional Neutral Network

With the rapid development of computer technology,especially the gradual life of artificial intelligence and pattern recognition in recent years,people's demand for intelligent machines is also daily on the increase.People not only expect robots to help humans finish physical work,but also expect robots to judge people's psychological activities and real intentions,so as to realize human-computer interaction process with the goal of emotional communication between humans and computers,which is helpful for better serving humans.In the process of communication,the face is the core part for human communication,which contains many sensory organs(eyes,ears,nose and mouth,etc.).Through facial communication,others' intentions and emotions can be understanded quickly in the most direct and natural way.At the same time,other preople' psychological activities and real intentions can be judged by recognizing other's facial expression.Therefore,how to make robots quickly and accurately recognize human facial expressions has become a major focus of intelligent human-computer interaction technology research.As an intelligent human-computer interaction technology,facial expression recognition is an important step in the field of artificial intelligence,which has been widely concerned and studied by many scholars.At present,the performance of image classifier largely depends on whether the extracted features are effective or not.The convolutional neural network model has a significant advantage in the field of computer vision and artificial intelligence.It is of great theoretical significance and practical value to study facial expression recognition through convolutional neural network model.This paper focuses on improving the accuracy and computational efficiency of facial expression recognition by using the convolutional neural network model.Firstly,a data augmentation method is proposed to obtain high-quality images for the training of convolutional neural network model.Then,a face detection method to reduce the misjudgment rate is proposed based on the acquired big data set.At the same time,the corresponding convolutional neural network classifiers are designed based on the features of the facial expression images,so as to realize the accurate recognition of different facial expressions.This dissertation mainly includes the following aspects:1.A face detection method to obtain expression region in complex background is presented in this paper.Because most facial expression images have complex background information,which will affect the effective extraction of facial expression feature information,and it is not conducive to the final expression recognition,an effective face detection algorithm is proposed in this paper.The algorithm consists of two aspects.Firstly,the traditional skin color model was used to detect the facial expression for the first time,and then used the method based on eye positioning to adjust the face image for the first time,which can reduce the false detection rate,and effectively reduce many interference factors in the face recognition task.At the same time,in order to speed up the speed of eye positioning,the image region segmentation method is adopted to reduce the time of eye searching and speed up the operation of the system.2.A generative adversarial network model is constructed in this paper to reduce the collapse of facial expressions.Because the training of the convolutional neural network model is inseparable from the support of large-scale data set samples,and most of the existing face expression data sets are generally small in size,a generative adversarial network model for data enhancement is proposed.The model consists of generator and discriminator based on neural network architecture.The architecture and objective function of the model are optimized.The advantage of this model is that it can reduce the generation of collapse images by increasing the reconstruction error,improve the image generation quality,and make preparations for the subsequent model training.3.An auxiliary convolutional neural network classifier based on the important components is proposed for some exaggerated expression images.Because most of the existing methods only use the whole image of human face as the input information,which will fail to grasp the key feature information that is beneficial to classification.Firstly,the original face image is used to obtain the feature information of the first layer,and then the feature information of the key areas is extracted,which is used to be fused with the feature information in the first layer.In addition,a new piecewise activation function is proposed to reduce the shortcoming of concussion in the training process.At the same time,a method based on CNN and random forest classifier is proposed for solving the problem of time-consuming problem.In order to improve the efficiency of the random forest classifier,the formula of the information gain rate is simplified,and the decision algorithm of the random forest classifier is optimized.4.In view of the confused samples,a cascade convolution neural network identification model is proposed in this paper.In addition,the greedy algorithm is used for fusing multiply low-dimensional feature informations,which can reduce the impact of the dimension disaster.At the same time,a model based on clustering and convolution neural network is proposed in this paper for largely confused expressions.In the process of clustering,a new clustering algorithm based on fixed initial values is proposed in order to increase the clustering center distance,which is helpful for improving the whole recognition rate and the expression of facial expression recognition rate under each category.5.A hybrid transfer algorithm based on the fusion of convolution restricted boltzmann machine and convolutional neural network model is proposed for some large scale facial expression database.Although the data augmentation method is beneficial to the training of the convolutional neural network model,the data augmentation processing process is also complex and the training time of the model is long for some large data sets.Therefore,this chapter proposes to apply thetransfer method to a large data set.This algorithm model finish feature extraction of the source domain for the first time by using convolutional neural network model,and then uses the target domain to continuely learn more advantageous characteristic information on the basis of the first learning features;Secondly,using convolution restricted boltzmann machine to acomplish the deep excavation for the target domain knowledge,which can decrease the content differences of the data set in the process of transfer algorithm.In addition,in order to make the convolution restricted boltzmann machine obtain more comprehensive edge feature information in the convolution operation of the visible layer,the method of zero complement operation is adopted.