| Expression information can convey a person’s emotional state and intention of action,which is the main way of daily communication.With the advent of artificial intelligence and the era of big data,Facial expression recognition has become a hot research topic in the fields of pattern recognition and machine vision.Facial expressions are a unique way for humans to communicate with each other.The ability to give the robot a sense of facial expression is a key technology for intelligent human-computer interaction.With the development of computer hardware and software technologies such as GPU and memory,facial expression recognition technology is more widely used,but it also puts forward higher requirements for facial expression recognition technology.The traditional facial expression recognition method mainly relies on artificially designing a feature extraction algorithm suitable for a certain scene.In this way,the feature extraction steps in the early stage are cumbersome,the calculation process is complicated,the feature dimension is huge,and the expression of valuable facial expression information is also affected.The traditional way of identifying facial expressions is difficult to meet the actual needs,which has great limitations.Deep neural network relies on pure data,breaking the way of complex feature extraction,which can implicitly learn more valuable expression feature information.Face expression images obtained by face detection in real life tend to have partial occlusions,which are mainly caused by external environments such as limb movements.The occluded face area often contains a large amount of valuable expression feature information,and the feature extraction is directly performed,which may cause the expression feature information to be insufficiently expressed,which directly affects the recognition result.This paper proposes an improved convolutional neural network for facial expression recognition.The paper mainly completed the following work:(1)The theoretical algorithm of deep learning is studied.In view of the current mainstream expression datasets from abroad,in order to eliminate the impact of ethnic differences,the datasets collected from real scenes were established and verified.Data processing such as face detection,grayscale,equalization and data enhancement was performed on the Fer2013 and CK+ datasets to eliminate the interference of external factors on feature extraction.Based on the Tensor Flow deep learning framework,a convolutional neural network with continuous convolution characteristics is built.The network has a14-layer structure,and uses the cross entropy loss function to calculate the loss.The Adam optimizer is used to optimize the weight parameters.It is iteratively trained on the Fer2013 and CK+ facial expression data sets,and the accuracy rate reaches 94.12.% and 93.75% are better than other algorithms.(2)The deep convolution generation is constructed to defend the occlusion facial expression image against the network,because the restored facial expression image can not represent the feature information of the real image,and the expression recognition effect is not ideal.Based on this problem,a two-channel convolutional neural network is constructed.The convolutional neural network refers to the Inception module and the skip connection structure.It has a two-channel structure,the main channel convolutional neural network is used to train normal facial expression data sets,the auxiliary task channel adopts the idea of multi-task learning,and the restored occlusion facial expression image is used as a training data set to learn its Characteristic information,the two features are merged,the loss calculation is performed by using the cross entropy loss function,and the weight parameters are optimized by the Adam optimizer.The training is performed on the Fer2013 and CK+facial expression data sets,and the accuracy rates are respectively reached.88.17% and86.76%,the recognition effect of partially occluded facial expression images is stable. |
PDF Full Text Request