100 Neural Networks as a Useful Tool for Real-Time Facial Expression Recognition
Download citation file:
Human beings communicate by rather skillful projection and reception of facial and voiced-expressions, as well as gestures (hand, posture, etc.). In order to develop “tools” by which humans can interact with a computer, we need to develop a software tool that is capable of processing, for example a wide range of facial expressions (FEs) in time. Facial expressions, in their variety and in addition, digital quality when captured presents itself as nontrivial challenge in applied biometrics. In this paper we present results to date on development of a real time and computationally “light” facial expression recognition software tool. Contrary to past studies reporting elaborate processing and FE classification methods, we undertook an approach extracting a small number of facial feature points, and one that realistically contained noise. We additionally proposed a vectoral descriptor and amplitude for a given FE, relative to a reference, neutral FE image. Moreover we trained and tested across FE databases that were ethnically∕culturally and gender-wise different. Our results to date opened for consideration the following in brief, that: 1) there are similarities∕differences across FE database, 2) vectoral descriptors and amplitude per FE appear effective, 3) “happy” and “surprise” FEs are well classified across different database, while 4) “angry-distress” and “fearsurprise” pairs are linked by misclassification across databases; that is, there is some evidence that these are ethnicity∕culture specific and therefore misinterpreted.