Research On Facial Expression Recognition Based On Convolutional Neural Network

Facial expression plays an important role in daily life.The information transmitted through the facial expression is far more than language and speech.Nowadays,humancomputer interaction is becoming more and more frequent.Understanding human emotions and interacting with people naturally are also the inevitable direction for the development of the future robots.Moreover,as a key technology in affective computing,facial expression recognition has broad application prospects in the fields of pension,education,medical treatment,and entertainment.Facial expression recognition is an important and challenging research topic in the computer vision field.The differences caused by changes in facial expressions are very subtle and the differences between different expression features are small,which makes it difficult for computers to accurately recognize the facial expressions.In addition,the input of expression recognition networks is often low-resolution.But many common convolutional neural networks need large size images as the input.Excessively upsampling the image to fit the network not only increases the calculation but reduces the recognition effect.Aiming at these problems,this paper studies the perfection of convolutional neural network and loss function.The main work and contributions of this paper are as follows:1.A novel dilated convolutional neural network for low-resolution images is proposed.Based on the dilated convolution operation,the proposed method divides the network into front and back two parts,which can reduce the use of pooling operations in the model and is better for low-resolution images.This design deepens the network as well.In addition,the Jagged dilate ratio design guarantees the size of the receptive field in the feature maps and improves the recognition accuracy.2.A joint loss function based on the center balance loss is proposed.The models trained with center balance loss function can have better feature extraction capability.Considering the characteristics of the facial expression recognition dataset,a balance term is introduced to the computational process based on the center loss function when the category center is updated.This way can restrict the distance between different categories centers.The joint loss function of cross entropy and center balance loss is adopted for model training to improve the recognition effect.3.Based on the convolutional neural network proposed in this paper,a facial expression recognition method for video image sequences is designed.With the structure of convolutional neural network and long short-term memory network,the proposed static image-based model is extended for facial expression recognition of video image sequences.