The Research On Micro-expression Collection And Processing Method Based On Crowdsourcing Data Cleaning

Micro-expressions are quick expressions of people that hide their true emotions.In recent years,due to the potential applications of micro-expressions in various fields such as mental health,national security,and polygraph detection,the research of micro-expressions has attracted numerous attention of more and more experts in different scientific areas.Currently,the research on micro-expression has made some progress.Supervised learning methods such as deep learning-based models are widely adopted in micro-expression detection and recognition,which usually require a large amount of labeled micro-expression data for training.However,there is still a lack of high-quality micro-expression datasets,which seriously hinders the development of micro-expression research.Most of the existing micro-expression data sets are collected in a laboratory environment.Most of this collection method is time-consuming and labor-intensive,and only a small number of samples of a single category can be obtained,which is difficult to meet the needs of deep learning model training.At present,crowdsourcing has been widely used in data collection in other fields due to its collective intelligence synergy.Therefore,this paper proposes two crowdsourcing-based micro-expression collection and processing methods,which can collect a large number and variety of samples,and can detect and label micro-expression data more efficiently.Due to various unstable factors of crowdsourced labelers,the quality of data labeling will become low.This thesis ensures the quality of the data by cleaning the crowdsourced redundant data.(1)This thesis proposes a crowdsourcing-based method for collecting and processing multi-source micro-expressions.In terms of current micro-expression collection,most of the public micro-expression data are collected in a laboratory environment,and the number of collected micro-expression samples is small and unnatural.Therefore,in this thesis,the preliminary collection of micro-expressions is not limited to the laboratory environment,and the preliminary collection of micro-expression samples is carried out through various methods such as Internet data resources,related data sets,system web pages,and We Chat mini-programs.In this thesis,the micro-expression video clips are initially detected by the main direction maximum difference analysis method,and then the micro-expression detection and labeling are carried out by crowdsourcing.Finally,the data is integrated to obtain the final micro-expression dataset.Compared with other micro-expression collection methods,the method can collect more diverse and natural micro-expression data more efficiently.(2)This thesis proposes an efficient,crowdsourcing-based large-scale micro-expression collection and processing method.In order to obtain large-scale natural micro-expression samples,this paper firstly collects long video clips containing micro-expressions on a large scale,and then uses the micro-expression detection method based on optical flow and long short-term memory to detect the long video clips containing micro-expressions obtained above.Further data processing is performed to obtain a micro-expression video.The micro-expression videos are preliminarily screened,and the screened micro-expression videos are placed on the Amazon Mechanical Turk for annotation.The labeled micro-expressions are cleaned by the DS algorithm to obtain the final micro-expression dataset Crowd4ME2.Compared with other micro-expression datasets,our micro-expression data has obvious advantages in quantity and collection efficiency.Based on the above methods,this paper futther designs a micro-expression collection and management system with crowdsourcing,which integrates micro-expression data collection,labeling and management,facilitating the collection and processing of micro-expression data.