Research On Improved Generative-Adversarial-Network Based On Micro-expression Recognition Algorithm

Micro-expression is a kind of spontaneous facial expression with short duration,small amplitude,and small action area.Micro-expression recognition plays an import role in psychological diagnosis,case detection,danger warning,public safety,business negotiation and foreign affairsAt present,the biggest difficulties in the micro-expression recognition are subtle differences among samples and lack of sufficient training samples which lead to low recognition rates.There are currently three types of methods for improving the recognition rate,one of which is mainly a traditional method,such as a method for extracting fine and subtle expression features.One is mainly deep learning methods,such as trying to transfer knowledge from other fields to improve recognition performance.Another is to try to generate samples using the generated model,and add the generated samples to the training set to enhance the sample recognition accuracy However,due to the lack of research on micro-expression generation at home and abroad,it is rarely used.In view of the above situation,this paper proposes two kinds of micro expression generation methods:the micro expression generation model based on image,action and optical flow decomposition and the micro expression generation model based on expression,content decomposition and reconstruction.Specifically,the main contributions of this paper are as follows:·We propose an adversarial network model based on multi-labels micro-expressions of action units.Facial action units(Action Units,AU)are added to generate adversarial networks in the form of multi-labels.The time series and labels of micro-expressions are generated by the time series generator and the video generator.The image generator makes micro-expression images clear and video action effects good.The image discriminator ensures the generated image content highly similar to real micro-expressions.The video discriminator conforms the action of the generated content to the fine action trajectory of facial action units The optical flow discriminator ensures the subtle differences in facial movements A simplified attention mechanism ensures the quantity of high-quality generated micro-expressions.By adding the generated samples as data-augmentation samples to the training data set of micro-expression recognition,the recognition accuracy rate is improved by 7.4%,4.3%and 7%from CASME ?,MMEW and SDU,which proves that the generated micro-expression samples are of high quality and strong practicality.·We propose a method for generating micro-expression sequences based on expression,content decomposition and reconstruction.This algorithm uses facial expressions and action information of macro expressions.It adopts the idea of cross-domain.The identity information of the macro expression can not only expand the diversity of the micro expression database,but also maintain the identity information of the generated micro expression.The reconstructed thought guarantees the authenticity of the generated macro and micro expressions.By adding the generated samples as data-augmentation samples to the training data set of micro-expression recognition,the recognition accuracy rate is increased by up to 7.4%,6.4%and 10%from CASME ?,MMEW and SDU,which proves that the generated micro-expression samples are of high quality and strong practicality.