Facial Expression Recognition Algorithm Based On Deep Learning

Facial expression recognition algorithm is an important part of face recognition technology.It is widely used in the fields of human-computer interaction,security monitoring,autonomous driving and so on,and has become a research hotspot in academic and industrial fields in recent years.The facial expression recognition algorithm performs modeling based on facial expression features,so that the model can recognize specific expressions in the image or video,determine the psychological emotion of the object,and then perform related processing on the object.Existing facial expression models are difficult to capture the subtle movements of facial muscles,and are easily affect by factors such as lighting changes,occlusion,non-positive head posture,and identity information.In view of the above problems,A static facial expression recognition algorithm based on Alex Net-Emotion and a dynamic facial expression recognition algorithm based on attention mechanism are put forward in this thesis.Experiments demonstrate that both algorithms achieve high accuracy and robustness in static and dynamic facial expression recognition scenario,respectively.The research contents and innovations of this thesis are summarized as follows:(1)In order to solve the problem that it is difficult to accurately extract the expression features in static expression recognition scenario,a static facial expression recognition algorithm based on Alex Net-Emotion is proposed.It uses a smaller convolution kernel to improve the Alex Net network to reduce the model parameters and effectively detect the fine movements of facial muscles.At the same time,aiming at the problem of model over fitting caused by the small training data set,the batch normalization layer and PRe LU nonlinear activation function are added to accelerate the model convergence and alleviate the over fitting phenomenon.In addition,the angle change feature of key points of the face is extracted as auxiliary information to make the model maintain identity invariance and enhance the expression capabilities of facial expression features.(2)Aiming at the problem of small inter-class variance and large intra-class variance caused by only using the softmax loss,a joint optimization method of softmax loss and improved Island loss is proposed.When training the network,the two task loss functions exert their respective advantages to make the samples in the cluster as compact as possible and keep the clusters as far away as possible,so as to improve the discrimination power of the model.In order to solve the problem that Island loss cannot effectively distinguish between easy samples and difficult samples when calculating the distance between the samples and the corresponding class center,the online hard sample mining strategy is adopted to calculate a part of samples which exceed the radius ? of the center point,so that the model can focus on the more difficult samples.Therefore,the speed of model convergence can be accelerated,and the classification ability can also be improved.(3)As the static expression recognition algorithm is difficult to effectively identify the facial expression in dynamic scenes,a dynamic facial expression recognition algorithm based on the attention mechanism is proposed in this thesis based on the Alex Net-Emotion network.The algorithm consists of three modules: feature embedding module,recurrent neural network and attention module.The feature embedding module is composed of Alex Net-Emotion network,which is used to accurately extract the expression feature of each frame.The recurrent neural network is used to model the temporal data so that the model can effectively use the context information of the current frame to generate a hidden layer state feature representation of each frame.In order to distinguish the importance of video frames with different expression strengths,an attention module is introduced to calculate the weight of hidden layer state features at each moment,and linearly weighted to generate video-level expression feature vectors,which effectively solves the problem that it is difficult to effectively integrate the expression features of video frames in dynamic scenes.