Facial Expression Recognition Using Learning Vector Quantization

Although the detection of emotions from facial video or images has been topic of intense research for several years, the set of applied classification techniques seems limited to a few popular methods. Benchmark datasets facilitate direct comparison of methods. We used one such dataset, the Cohn-Kanade database, to build classifiers for facial expression recognition based upon Local Binary Patterns (LBP) features. We are interested in the application of Learning Vector Quantization (LVQ) classifiers to this classification task. These prototype-based classifiers allow to inspect of prototypical features of the emotion classes, are conceptually intuitive and quick to train. For comparison we also consider Support Vector Machine (SVM) and observe that LVQ performances exceed those reported in literature for methods based upon LBP features and are amongst the overall top performing methods. Most prominent features were found to originate, primarily, from the mouth region and eye regions. Finally, we explored the specific LBP features that were found most influential within these regions.