Font Size: a A A

Research On Micro-expression Recognition Technology Based On Deep Learning

Posted on:2019-09-23Degree:MasterType:Thesis
Country:ChinaCandidate:J H PangFull Text:PDF
GTID:2428330542999666Subject:Information and Communication Engineering
Abstract/Summary:PDF Full Text Request
With the improvement of graphics ability and the advent of the age of big data,the research direction of pattern recognition under computer vision has been developing steadily.Expression recognition and micro-expression recognition have been widely applied in many fields.Micro-expression recognition technology has been paid more and more attention in recent years as a kind of new recognition technologies,however,the development of the recognition technology is faced with the challenges of low recognition rate and insufficient samples because of the particularity of micro-expressionsAt present,the difficulty of micro-expression recognition is the constraint of the short duration of micro-expressions,small changes,lack of adequate training samples,etc.,however,its application needs in lie detection and psychological prediction are not in line with the current research results.To solve the problem of micro-expression recognition,this paper proposes a micro-expression recognition technique based on deep learning,specifically,the main contributions are summarized as follows: This paper describes the largest micro-expression database,SDU2,published by our research team,and a method of alignment and cutting of microexpression database is proposed. This paper presents a method of micro-expression recognition based on short-and long-term memory networks,which aims to solve the difficulty in deep learning of micro-expressions' time characteristics,three databases,CASMEI,CASMEII and SDU2,were used to verify the validity of the feature extraction method of deep learning. In order to increase the number of training samples,as positive and negative samples,the macro expression is used to train the micro-expression.Secondly,to retain the local information of micro-expressions while extracting global features,a 'macro to micro-transformation model' is proposed.Finally,to distinguish the micro expression from the macro expression to extract the characteristic of micro-expressions,a cross-modal triplet loss function is proposed to train the network.The experimental results show that the 'macro to micro-transformation model' is superior to other modern methods.
Keywords/Search Tags:Micro-expression recognition, Deep learning, Cross modal, Convolutional neural network, Long short-term memory
PDF Full Text Request
Related items