Face continuous expression recognition method based on time sequence attention mechanism

Abstract

The invention relates to a face continuous expression recognition method based on a time sequence attention mechanism, which is characterized in that video frame feature extraction is carried out through a deep convolutional neural network, a space-time attention mechanism is combined, the description capability of features in a space domain is enhanced, the description of the features in a time domain is increased, and in a continuous-dimension emotion space, and sentiment prediction related to the spatio-temporal context is performed. Emotion change is a progressive process, and good continuous emotion recognition precision is difficult to obtain only through expression features of a single-frame spatial domain. The image frames close to the time domain generally have regularity, and time domain feature calculation can provide reliable data support for learning-based multi-frame fusion features. The method comprises the following steps: extracting a context dependency relationship of continuous multi-frame expressions in a video based on a pleasure-activation emotion space and a space-time attention mechanism, and learning a facial muscle movement rule in an expression generation process; and establishing a continuous expression recognition model. The method can be applied to the fields of criminal investigation, civil aviation safety detection and the like.

Description

technical field [0001] The invention relates to the field of artificial intelligence and computer human-computer interaction, in particular to a video-based continuous dimension emotion recognition method. Background technique [0002] With the deepening of artificial intelligence and emotional psychology research, it is possible to use computers to automatically identify and monitor facial micro-expressions in videos, and analyze emotional state behaviors. Moreover, emotional analysis of human faces in videos is widely used in criminal investigation, civil aviation safety testing, education and training. , special medical treatment and e-commerce have high application value. The emotional analysis of the suspect or the monitored person is helpful to assist criminal investigators to correctly identify the relationship between behavior and psychological emotion, and provide a basis for judgment. In the inquiry of the transportation of dangerous goods in civil aviation or the...

AI Technical Summary

Problems solved by technology

[0003] Aiming at this problem, the present invention proposes a facial continuous expression recognition method based on a temporal attention mechanism, which extracts the spatiotemporal salient features of the human face, improves the recognition accuracy, and solves the problem that the dependency relationship between frames cannot be accurately expressed, the training is unstable and the The problem of large output value jitter, the present invention also provides an improved 3D convolutional neural network model, the model has added time and space attention layer, the purpose is to better represent the dependence between the space-time position; the model uses The expected regression layer replaces the traditional single neuron regression layer, the purpose is to solve the problem of gradient instability and output jitter during the training process

Images