Research On Facial Expression Recognition Algorithm Based On Deep Learning

Human facial expressions are the expression of inner emotions and an important means of communicating emotions between people.With the development of artificial intelligence,facial expression recognition has been widely studied as an important topic in computer vision,and has important applications in the fields of human-computer interaction,online learning,medical diagnosis and intelligent driving.Therefore,how to obtain a model with higher facial expression recognition rate and faster recognition with fewer parameters is very important for the facial expression recognition task.Based on different feature extraction methods,this paper mainly proposes four improved expression recognition algorithms.First of all,early manual facial expression feature extracting algorithm has certain effect on the small data set,such as local direction number(LDN)algorithm,but only when the algorithm expression features are extracted to extract the edge response is encoded in the direction of maximum and minimum,ignoring the other larger image edge response direction contains information.To solve the above problems,an improved local directional number pattern(ILDN)algorithm was proposed,which encoded the directions of the maximum value and the sub-maximum value of the absolute value of edge response and the symbols of the two values.The statistical histogram encoded by 8-bit binary was used as the facial expression feature.The recognition rate is 92.43% on JAFFE dataset and 94.5% on CK+ dataset,11.7% higher than LDN algorithm.By local orientation mode(LDP)and its improved algorithm,the spatial local orientation is put forward several modes(DSLDN)algorithm,and space in the direction of the maximum value space and difference encoding direction information,response value on the edge of the gradient space for sorting,before taking a assignment to 1,the rest is 0,with strength of eight binary coding information.The direction information and intensity information were extracted from the histogram and then cascaded as facial expression features.The recognition rate is 93.88% on CK+ dataset,which is 11.08% higher than LDN.Secondly,with the development of deep learning,various deep convolutional neural networks such as VGG have been produced in the field of image recognition,and the idea of transfer learning has also achieved good results in facial expression recognition tasks.However,for small data sets,the over-extraction of facial expression image features by deep convolutional neural network greatly improves the probability of over-fitting the model,and the over-deep model with more parameters will lead to the waste of computing resources.To solve this problem,a shallow dense convolutional network(SDN)was designed by using the special feature connection mode in the dense convolutional network.The number of convolution kernels in the feature extraction module of the dense network model was increased to increase the model width and reduce the model depth and parameter number.The shallow dense convolutional network(SDN3)achieves 95.25% and 94%recognition rates on CK+ dataset and JAFFE dataset.Finally,the manual feature extraction algorithm may lose the details of facial expression images in the process of coding,and the dimensionality reduction operation in the feature extraction process of convolutional neural network may also lose part of facial expression images.Aiming at the above problems,a two-channel parallel network model is constructed.A two-space local directional number model was used as one channel to extract facial features,and a shallow dense convolutional network was used as the other channel to extract facial features.The two features were fused and then classified.The algorithm achieves 97.45% and 96.95% recognition rates on CK+ dataset and JAFFE dataset.