Research And Implementation Of Expression Recognition System Based On Lightweight Neural Network

With the rapid development of computer technology and artificial intelligence,facial expression recognition technology has become a current research hotspot,and has broad application prospects in many fields such as human-computer interaction and safe driving.The traditional facial expression recognition technology is not only difficult to design,but also cannot fully extract facial expression features,resulting in poor facial expression recognition effect.At present,the expression recognition technology based on deep learning has become the mainstream,but the current convolutional neural network model has high hardware requirements and takes too long to train,which makes it difficult to guarantee the real-time performance of expression recognition and cannot be deployed on devices with low computing power.Aiming at further improving the accuracy and robustness of expression recognition algorithm and reducing the parameters and calculations of convolutional neural network,this paper studies the lightweight method of convolutional neural network and expression recognition method,and realizes a set of facial expression recognition system.The main research contents of the paper are as follows:First of all,in the original Ghost Net model,some neurons do not work and cause information loss in the Ghost bottleneck layer during model training.The Mish activation function is used to design the improved Ghost bottleneck layer.For the SE attention used in the original Ghost Net model,the module does not take into account the information at the spatial level,and uses a more efficient and lightweight Normalized attention module as an alternative.Aiming at the problem that the original Ghost Net input scale is too large,the initial input size and layers of Ghost Net model are modified to reduce many parameters and calculation,so as to make it more suitable for expression recognition task.Based on the above three points,a modified Ghost Net model is proposed.Experiments are carried out on the FERplus and CK+ data set,and the accuracy rates is 79.339%and 96.411%,respectively.Experiments show that compared with other mainstream lightweight convolutional neural networks,the method based on the improved Ghost Net model still has higher recognition accuracy and faster recognition speed on the premise of less model parameters and calculation.Secondly,the inter-class differences of facial expressions are small,and the intra-class differences are large.The Softmax cross-entropy loss function cannot compress the intra-class space,and the optimization is not flexible.Based on the above problems,combined with the design ideas of Island loss function and Circle loss function,a loss function based on cosine similarity is designed and adopted to supervise the training and learning of neural network.The method can reduce the intra-class difference and increase the inter-class difference in the feature space,thereby improving the feature discrimination ability.Experiments on FERplus and CK+ data set shows that the accuracy of loss function based on cosine similarity is as high as 82.561% and 98.553%,which is better than softmax cross entropy loss function,island loss function and loss function based on cosine distance.It is more suitable for the use scenarios of mobile terminals and embedded devices.Finally,in order to apply the trained lightweight convolutional neural network in practice,Py Qt5 tool package is used to complete the design of the whole facial expression recognition system.The system has the following functions: static expression recognition through image recognition,dynamic expression recognition through camera recognition,display the time required for recognition and recognition results,and display the expression prediction probability distribution in real time.The system test shows that the system has good real-time performance,and has an accuracy of up to 90% on the positive face.