A main goal in the field of affective computing is to train machines to sense and recognize emotions from image data using techniques from Computer Vision. Of interest to this study is the recognition of micro-expressions: short and involuntary expressions of emotions which cannot be concealed, revealing reflexive or reactionary emotions. The study of micro-expressions using Computer Vision methods is an emerging field for which methodologies and datasets are newly being contributed addressing the short and involuntary characteristics of this type of facial movement. Challenges in the automatic recognition of microexpressions include the spontaneous nature which needs to be present in data collected, as well as the short duration of expression. Early research has addressed the challenges of micro-expression detection by using very complicated methods during the feature extraction phase (such as temporally interpolating frames to add resolution in time) or highly sophisticated machine learning algorithms (such as the Gentleboost ensemble methods). Previously proposed methodologies have proven difficult to implement for replication. This paper presents an approach that uses simpler constructs than proposed in the literature for both image representation (feature extraction) and classification (machine learning) while maintaining better performance than the published baseline for the given dataset. Using Gabor features extracted along 9 scales and 8 orientations on images sub-sampled to 14 × 11 pixels and multi-class Decision Trees, we were able to achieve an accuracy of 81.7%. This outperforms the baseline accuracy of 63.4% obtained using Local Binary Pattern on Three Orthogonal Planes (LBP-TOP) features and Support Vector Machine classification.
展开▼
