Research On Micro-expression Recognition Based On Subspace Learning

Micro-Expression(ME)is one type of brief,subtle,and involuntary facial expressions,whose duration is often within half a second.It can expose the genuine emotions which people try to conceal.The aim of micro-expression recognition(MER)is to enable the machine to understand these genuine emotions.Based on MER,we can develop lots of interesting and useful techniques,e.g.,lie detection.Due to this fact,recently MER has been one of the most attractive research topics in affective computing,pattern recognition and computer vision.In this dissertation,we focus on two challenging problems in MER,i.e.,(1)how to locate the ME-aware facial local regions and(2)cross-database micro-expression recognition(CDMER)problem,in which the training and testing belong to two different ME datasets with large differences,and propose a series of effective and well-performing subspace learning methods.Concretely,this dissertation has following four main contributions:1.Firstly,we propose a novel MER method consisting of a hierarchical spatial division scheme and a kernelized group sparse learning(KGSL)model,which takes full advantage of the facial local regions associated with MEs.In this method,the hierarchical spatial division scheme aims to produce such contributive facial local regions and thus spatiotemporal descriptors extracted based on the hierarchical scheme(hierarchical spatiotemporal descriptors)can be used to better describe MEs.Meanwhile,by using KGSL to process hierarchical spatiotemporal descriptors,we can pick out the elements contributing to MER from all the facial local regions yielded by the hierarchical scheme and further measure their detailed contributions.2.Secondly,we propose a simple yet effective domain adaptation method called target sample re-generator(TSRG)to deal with CDMER problem.TSRG targets at learning a sample regenerator to regenerate the source(training)and target(testing)ME samples.After that,the regenerated source ME samples keep unchanged in the feature space,while the regenerated target samples are enforced to share the same or similar feature distributions with source samples.Thus,we can use the classifier trained based on labeled source ME samples to predict the ME categories of the target samples.By the way,to the best of our knowledge,this is the first work to investigate CDMER problem.3.We further extend TSRG proposed in Contribution#2 to a generalized framework called domain regeneration(DR),which inherits the basic idea of TSRG,to deal with CDMER tasks.The subspace serving as the regeneration platform in DR framework is extended to the arbitrary one instead of the original feature space used in TSRG.Under the DR framework,we design three different domain regenerators including Domain Regeneration in the original Feature Space with unchanged Source domain(DRFS-S)and Domain Regeneration in the original Feature Space with unchanged Target domain(DRFS-T)whose is regeneration platform the original feature space and Domain Regeneration in the original Label Space(DRLS)with the label space as the regeneration platform.It is worthy to mention that TSRG proposed in Contribution#2 can be actually interpreted as one of the above three domain regenerators under the DR framework,i.e.,DRFS-S.4.We propose a novel CDEMR method consisting of an auxiliary set selection model(ASSM)and a transductive transfer regression model(TTRM).In this method,ASSM is firstly used to select a few samples from the target ME database to serve as the auxiliary set samples and then together with source ME samples to train the TTRM such that the feature distribution difference between the source and target samples can be relieved.Different from TSRG and DR framework,TTRM borrows the basic idea in Contribution#1 and further considers the different contributions of facial local regions to CDMER.