Research On Facial Expression Recognition With Categorical Model And Expression Ambiguity

Facial expression recognition has become one of the most popular directions in the field of the computer vision,and has been applicated in the case of driving assistant,intelligent class,and customs inspection,etc..Based on discrete categorical annotations,facial expression recognition is mainly achieved using deep learning to extract low-level features,and analyze the affect of the human face.However,most facial expression recognition algorithms have suffered from ambiguity problem due to the subjective recognition on emotion and the complicated mechanism of the expression.The ambiguity problem will make semantic information lost,worsen the algorithm performance,and cause a bias between model parameter distribution and real scene.Thus,analyzing and solving emotion ambiguity has great significance on expression recognition,and need to be solved as soon as possible.This paper mainly researched the facial expression recognition model design and facial expression ambiguity problem.The works and related innovations are:1.Propose an algorithm of facial expression recognition with high accuracy.Analyze and compare the similarities and differences between facial expression recognition and traditional image classification,and establish an expression dataset based on the difficulty of facial expression recognition.The dataset contains many facial images collected from the Internet,and is various enough to cover the existing expression types.Then A deep network for facial expression recognition is designed.In addition,class-weighted sampling is adopted for data imbalance.Through lots of ablation studies,the general metrics of facial expression recognition model design are proposed.2.Propose a method of expression ambiguity modeling and applications.For the ambiguity problem in facial expression recognition,a quantitative metric for expression ambiguity evaluation was proposed first.Then,an ambiguity measurement model is proposed to make it possible to accurately judge the degree of ambiguity from facial expression images.The experiments also showed that the output of the ambiguity measurement model is consistent with people's subjective cognition.Finally,the extents of ambiguity are applied to the facial expression recognition network,which significantly improved the accuracy.This work basically completed the research of 2-D image facial expression recognition and expression ambiguity.Finally the automatic facial expression recognition based on computer is realized and the automatic measurement of facial expression ambiguity is achieved.It lay the first stone for the subsequent research on facial expression recognition.