Research On Key Technologies Of Micro-expression Recognition

Facial expression is an important form of information transmission between individuals.Over the years,focused on the modeling and analysis of expressions,affective computing is aimed at automatic tasks such as facial expression recognition,assisted face recognition,age estimation,and so on.Spontaneous expression,which is uncontrollable for humans,is gradually becoming a new research hotspot as a reflection of people's real emotions.As a type of spontaneous expression,micro-expression is involuntary.Unable to be hidden nor forged,it possesses great value in the field of clinical diagnosis and safety precautions.However,the automatic recognition of micro-expressions is very challenging for its short duration,low intensity and limited distribution on human face.Additionally,due to the great difference between micro-expressions and the normal ones and the deficiency of mature micro-expression recognition methods in the field of computer vision,a lot of additional works are required.At present,micro-expressions recognition has not yet been explored extensively.This paper aims to improve the effectiveness of automatic recognition of micro-expressions from four aspects as follows:1)Reviewing the literature and summarizing existing the researches on micro-expressions,including the established micro-expression databases and the development of micro-expression recognition.2)Inspired by the spatiotemporal properties of micro-expressions,a micro-expression recognition method based on cross pattern coding on three orthogonal planes is proposed.First of all,in order to correct rotation and translation,we have proposed a face-alignment method.Using partial face landmark points,the least-squares method is used to calculate the alignment matrix between two frames and all frames of a micro-expression are aligned to its first frame.According to the correspondence between Facial Action Coding System and muscles' movement regions,11 non-overlapping face blocks are produced.Feature extraction is performed in each block to avoid mutual interference.The complementary cross-patterns is used on three orthogonal planes to sample points and then encode them.Under the premise of maintaining the number of sampled points,multi-scale sampling is attempted to improve the ability of the operator to encode discriminative information.3)The nature of the micro-expression is a weak facial movement,so we propose a micro-expression recognition method based on optical flow and Fisher Vector encoding.Firstly,according to the vector superposition of the optical flow,we use mathematical statistics to estimate the optical flow of the irrelevant movement in the image as an offset for the original optical flow,and followed by the assumption of the local consistency of the micro-expression,the micro-expression sequence is divided along space and time.By using the Fisher Vector coding,we can obtain high-order information for the histogram of optical flow in each sub-block,in order to enhance its capability.4)Aiming at the problem of few standard micro-expression samples,the high cost of labeling by human labor,and the category imbalance existing in the dataset,a micro-expression recognition method based on active transfer learning is proposed.We try to use the united framework of active learning and transfer learning,on the one hand,we transfer label knowledge from semantically similar expression domain to the micro-expression domain,on the other hand,with this knowledge,active learning is used to query a small number of samples that contain the most information in the micro-expression domain.Those samples are labeled to train a better classifier.