Research On Cross-database Micro-expression Recognition Based On Domain Adaptive Method

Micro-expressions are produced by people when they intentionally or unintentionally conceal their true emotions.They are facial expressions that occur for a very short period of time,are weak and are not controlled by the nervous system.Microexpressions enable a more accurate understanding of mental states and emotions,and have unique advantages in communication and negotiation,marriage prediction,teaching and assessment,and justice,so they have received a lot of attention from researchers in recent years,and have achieved certain results.However,the training and test samples of existing micro-expression recognition studies are mostly from the same database,i.e.it is assumed that the training and test data obey the same distribution,but this assumption often does not exist in practical applications.Due to differences in external factors such as acquisition equipment,illumination,stimulus material,recording environment,etc.,the samples from different databases have different feature distributions,resulting in a dramatic degradation in the performance of traditional micro-expression recognition methods when applied in practice.This paper presents a series of studies based on domain adaptive methods around the challenging problem of cross-database micro-expression recognition.The main research components are as follows:(1)A new model algorithm based on a combination of subspace learning and joint distribution adaption is proposed for cross-database micro-expression recognition.The algorithm first uses subspace learning methods to map the source and target domains into their respective subspaces,and then introduces intra-class and inter-class distance to preserve the source domain discriminative information through linear discriminant analysis.Maximize the sample variance of the target domain and retain data attributes.Finally,the maximum mean difference is used to narrow the data difference between the two subspaces and reduce the distance between the two subspaces as much as possible to improve the recognition effect of micro-expressions.A large number of cross-database experiments were conducted in SMIC database and CASME II database.The experimental results show that the algorithm effectively improves the micro-expression recognition rate and enhances robustness compared with the current better-performing micro-expression recognition methods.(2)As most of the existing cross-database micro-expression recognition methods require model selection or hyperparameter tuning from which the better result is selected,a process that takes a lot of time and energy.To overcome this problem,the Intra-domain Structure Domain Adaptation algorithm is proposed.The algorithm mainly learns nonparametric transfer features through intra-domain alignment and uses intra-domain planning to learn a classifier to accomplish the cross-database micro-expression recognition task.Experimental results show that the proposed method is efficient and saves labor and time costs.Compared with other cross-database micro-expression recognition methods,the average recognition accuracy is up to 14.17%.