| Micro-expression is a brief facial expression characterized by short duration and low intensity.Although these expressions are constantly occurring on people’s faces,they were found to only last for 1/25 th of a second and impossible to hide.That is to say,due to the short duration of the micro-expression,most people do not notice them.Through a large number of psychological studies show that micro-expression is difficult to hide,it is the true representation of human emotion.Micro-expression recognition,detection and analysis have a wide range of applications in human-computer interaction,national security and polygraph testing.However,because of the characteristics of microexpression,it is difficult for a trained person to recognize the occurrence of the micro-expression,which limits the application of the micro-expression in the real scene.Because of the complexity of the micro-expression recognition is high and the reliability of the micro-expression recognition is low,the micro-expression detection is more valuable in the real-world application.In order to solve the above problems,this paper presents a method of detecting micro-expression based on deep learning features and directional features.Detection of spontaneous facial micro-expression is a challenging job.Microexpression detection is to detect the occurrence of micro-expression in a video.In this paper,a new facial landmarks detection method combined with directional features for detecting micro-expression pipeline is presented.Firstly,Facial landmark localization plays important roles in micro-expression detection.We employ a multi-task learning approach for training deep convolutional neural networks,multi-task learning is a kind of inductive transfer learning method.Its goal is to improve the generalization ability of machine learning model.The use of parallel training to learn multiple tasks in a way that makes the convolution neural network more accurate for the facial landmark detection.We investigated a deep convolutional neural networks architecture that is able to estimate the position of facial landmarks and separated the facial area into twelve sub-regions.Secondly,micro-expression is generated by the facial muscles,we extract the histograms of oriented optical flow features in each sub region,and the direction of micro-expression characteristics of each frame are normalized.Through the direction features can better explain the mechanism of micro-expression,normalized direction features make the calculation more efficient.Finally,we use the support vector machine to detect the micro-expression.Compared with previous work,our method has the following advantages: 1.Compared with previous works,our method has more accurate facial landmark localization.2.The generation of facial expression is the result of facial muscle movement.Based on the directional characteristics of optical flow features,the mechanism of micro-expression can better explain.3.Our method can be more efficient by normalizing the extracted directional features. |
