CNN micro-expression recognition method based on cavity convolution

Abstract

The invention discloses a CNN micro-expression recognition method based on cavity convolution. The method is based on a cavity convolutional neural network. On the basis of an MTCNN and a VGG16 network, automatic acquisition and recognition functions of face micro-expressions are completed. The network receptive field is effectively improved through hole convolution; loss of image edge features and fine features in the convolution process is avoided; the face micro-expression recognition algorithm which is larger in receptive field and higher in recognition accuracy has high robustness, meanwhile, in combination with face expression automatic recognition, a set of overall framework from face detection and discovery to micro-expression classification is formed, and manual screening is avoided to a certain extent by utilizing a similar network structure.

Description

technical field [0001] The invention belongs to the technical field of micro-expression recognition, and relates to a micro-expression recognition method, in particular to a CNN micro-expression recognition method based on atrous convolution. Background technique [0002] Micro-expression is a subtle change produced by the special muscle movements of the face. As a natural mechanism of facial behavior, it cannot be faked and reflects the true inner emotions of human beings. The cycle is maintained between 0.04-0.5 seconds. The range of movements is subtle and difficult to detect with the naked eye. It has wide application value in the fields of public security, psychological treatment, negotiation and communication prediction, etc. However, due to the limitation of professional training and time cost for manual identification, although there are professional tools, the effect of human identification is only 47%, which is difficult to carry out Mass promotion. [0003] Micro...

AI Technical Summary

Problems solved by technology

[0004] In terms of feature extraction, on the one hand, the features obtained by traditional feature extraction methods such as LBP operator and HOG feature points are usually shallow features of images, although these feature-based methods fully consider the changes in facial expressions from the perspective of spatiotemporal texture. , but it cannot effectively describe the structural information of the sample, and it is difficult to distinguish the relationship between high-dimensional features; Large-scale, unable to meet the real-time analysis requirements, and unable to complete large-scale micro-expression recognition; at the same time, there will be areas lacking sufficient gray level changes between consecutive frames of the actual shooting video, and there will be real motion missed detection

[0005] In terms of micro-expression classification, with the development of deep network, deep neural network began to be applied in the process of facial micro-expression recognition. Summarizing the existing micro-expression classification network, with the continuous increase of the number of layers, although sufficient feature information can be extracted , but it also increases the amount of calculation of the neural nodes, and key details will be lost in the convolution process; at the same time, the existing micro-expression databases are all clear face forward images obtained after video frame segmentation and screening in a specific environment. In the actual application scenario of the trained model, the recognition rate generally drops significantly compared with the test, and the application of the real-time system is not good.

Images