Research On Micro-expression Dataset Establishment And Cross-dataset Recognition Methods

Micro-expression is a special kind of facial expression,which is defined as certain subconscious facial action triggered by emotional or other psychological activities.Microexpressions cannot be consciously controlled by the individual,thus can reflect real mental state of the individual.Compared with common facial expressions,micro-expressions are characterized in the following aspects:their facial movements are minimal;only involve a few local facial areas;has extremely short duration.Due to these traits,micro-expressions are difficult to be observed by the naked eye.However,a small amount of "undetectable" microexpressions is sufficient to reveal people’s genuine emotions,which has important application significance in a variety of high-risk scenarios,such as criminal investigation,lie detection,psychological diagnosis,security and so on.In recent years,with the development of multimedia technology and pattern recognition algorithms,the field of micro-expression recognition has attracted more and more attention,where a large number of research results have been developed,greatly improving the recognition effect of micro-expression.However,there still exists various problems to be solved.First of all,micro-expression is very difficult to observe with the naked eye due to its small amplitude and short duration,which leads to huge cost of manual annotation and limits the amount of data of micro-expression datasets.Secondly,there is no standard or unified definition of micro-expression at the present,resulting in the deviation between different micro-expression dataset production procedure and their emotional categories.As a further impact,the micro-expression sample sizes of different categories vary greatly,whereas some categories are not included as common emotional categories in micro-expression recognition researches.To sum up,the most significant problems in the field of micro-expression recognition are the problem of small samples and uneven sample distributions.What’s more,the current micro-expression recognition algorithms generally only classify three categories,namely,positive,negative and surprise,and do not study the more detailed classification of micro-expressions.Based on the above problems,this paper established a micro-expression database to enrich the number of samples,and proposed two micro-expression recognition algorithms based on domain transfer to ease the problem of small samples by transferring information from other databases.Experiments carried out on six categories of emotion improve the effectiveness of the proposed methods.The contribution of this paper is elaborated as follows:Firstly,this paper establishes SDU micro-expression dataset to expand the number of micro-expression samples.The samples of SDU dataset were induced by stimulating videos and were recorded in controlled environment of the laboratory.Two experts annotated the emotion categories and facial action units with regard to the real emotional responses of the subjects.SDU micro-expression dataset includes 6 types of common emotions,featuring high resolution,high sample authenticity,large sample number,rich emotion categories and balanced sample distributions.SDU micro-expression dataset can greatly enrich the number of samples in the field of micro-expression research,which is of great significance for the development in future researches.Secondly,this paper proposes a micro-expression cross-dataset recognition method based on joint orthogonal non-negative tri-factorization with graph structure.After projecting samples from different domains,the common category information among features of different microexpression datasets is learnt through joint orthogonal non-negative tri-factorization.At the same time,the algorithm preserves the intra-and inter-domain structure information through the graph structure constraint,so as to retain the discriminant prior information.In addition,this algorithm uses real label information as the decomposed label matrix,which further improves the micro-expression recognition effect through supervised training.Thirdly,this paper proposes a cross-dataset recognition method of micro-expression based on cross-reconstruction with structural discriminative enhancement.The features of same categories from the two domains are projected and used as coefficient matrix to reconstruct each other,so that the features of two domains are fully close to each other.What’s more,the discriminant constraint is introduced to enhance the classification separability of the learned new feature representation,which further improves the recognition effect of multi-class microexpressions.