Özet:

Micro Expression (ME) is the leakage that occurs when people try to mask their involuntary and uncontrolled emotional response to an event in a risky environment. Because the person experiencing the emotion at risk tries to suppress it, its reflection on the face occurs in a low intensity, a specific region, and a very short time. Since the expression emerges involuntarily, it is not fake but completely natural. Thanks to the correct detection of these natural expressions, it can be used effectively in many fields such as forensics, clinical, and education. This study used preprocessing, feature extraction, feature selection, and classification tasks in the framework created for the ME recognition target. CASME-II, one of the literature's most widely used publicly available ME datasets, was used in the proposed framework. In the preprocessing stage, onset and apex Frames are taken from the image sequence of each video clip to be used in optical flow algorithms. These two frames obtained horizontal and vertical optical flow images of Farneback, TV-L1 Dual, and TV-L1. Then the features of these optical flow images were obtained using the convolutional neural network (CNN) model Xception and the traditional Gabor model. Recursive feature elimination with a cross-validation (RFECV) feature selection algorithm was used to filter the distinctive ones of these features. Finally, the SVC Linear classifier divided the filtered ME features into three classes: positive, negative, and surprise. The results obtained from the proposed ME framework showed an accuracy rate of 0.9248.

Anahtar Kelimeler: