Uma análise crítica sobre as técnicas 3D empregadas no reconhecimento de expressões faciais apoiada em revisão sistemática da literatura

Abstract

Digital three-dimensional (3D) images represent a non-invasive alternative for representation and automatic classification of human emotional expressions. 3D images capture spatial information about deformations on face surface and allow overcoming some restrictions frequently found in digital bidimensional (2D) images. Solutions for the 3D Facial Expression Recognition (3D FER) problem combine, generally, computer vision and machine learning techniques. Investigations in this domain frequently rely on the use of databases made available by research groups or academic institutions, being a particular challenge the fact of those databases have a limited number of samples of few subjects performing facial expressions. Usually, preprocessing, feature extraction and selection procedures in addition to automatic classification algorithms are applied to this data but many challenges remain for the robust recognition of human emotional expressions. A critical analysis of multiple approaches and alternative techniques that have been proposed to overcome those challenges can enrich and support decision making on application development and identify gaps for further research. In this direction, it is presented in this dissertation the evaluation of 49 3D FER specialized studies by means of a Systematic Literature Review. Three major aspects are taken into consideration: preprocessing of captured data, face representation decisions and classification experiments. The investigation allowed identifying the preferential types of data and regions of interest for face representation. Also, it made it possible to note a trend towards the adoption of methods independent of a priori detection of characteristic points (keypoints) on face. The critical analysis also allowed highlighting and documenting some methodological vulnerabilities found in preprocessing procedures and in the design of classification experiments. All findings were carefully documented in order to contribute to the decision making and guidance of researchers in 3D REF