Abstract:
Automatic facial expression recognition is a prominent and challenging research interest
with usefulness in a variety of fields. It plays an important role in the fields of human
computer interaction, data-driven animation etc. Success of most facial image analysis
solutions depend on an effective facial feature representation. This thesis presents a novel
appearance-based facial feature, the Local Transitional Pattern (LTP). LTP can extract
robust facial feature from a face image that gives accurate and reliable recognition
performance for expression recognition. The LTP operator applied on a pixel finds the
monotonic intensity transition of neighboring pixels at different radii. The micro patterns
thus found is enhanced with spatial information by tiling the image and taking histogram
of each tile. The final feature vector is a collation of these histograms. This feature vector
is then employed to classify expressions with well known machine learning method:
Support Vector Machine (SVM). Cohn-Kanade expression database is used to conduct
experiments comparing LTP descriptor’s performance against other well known
appearance based feature descriptors. It shows that LTP descriptor has higher accuracy
than LBP and Gabor descriptors and it is also more robust against non monotonic
illumination.
