Construction And Algorithm Implementation Of AU Dataset Based On Depression Micro Expressions

In recent years,artificial intelligence has made rapid development in the medical field.For mental illnesses such as depression,the diagnosis and treatment of China and the world is not optimistic.The work of this paper is based on the general direction of combining artificial intelligence and depression diagnosis,and intends to lay the foundation for auxiliary diagnosis by establishing a data set and implementing related algorithms for facial behavior analysis,that is,AU detection.At present,there are problems in this field that hinder its development,such as the lack of available data and the lack of strict data annotation standards.Therefore,this paper proposes a set of rules and auditing standards for AU data labeling,and establishes a AU data set based on this which named AnDing AU Data Set.The data set contains outpatient video data of 62 patients with depression and completely saves picture and sound information.At the same time,the key small segments in the video are intercepted and framed,and the AU intensity is manually labeled frame by frame.Later,this article designed experiments to verify the validity of the data set.The main works of the paper are as the following parts:First,the existing construction methods and experimental methods of large AU datasets are studied.While learning their key ideas,the experimental networks and methods based on the convolutional neural network are determined.Second,the thesis builds an AU dataset based on depression micro-expressions,including the source and collection of data,the labeling rules of the data,the design and implementation of the labeling tools,the labeling and review of the data,etc.,finally finishing more than 97,000 available image data for AU detection training.Third,after theoretical research on the convolutional neural network,the experimental design of the network structure including the region of interest and the long-short-term and memory,and a comparative experiment on the data set,using F1-score as the evaluation index to analyze the experimental results.The final experimental results verify the effectiveness of the stable AU data set constructed in this paper in the AU detection task and the effectiveness of the experimental algorithm network.