Evaluation of Edge Orientation Histograms in smile detection

Abstract

Smile detection received a enormous attention due to its famous application as a `smile shutter' in digital cameras. Edge Orientation Histograms (EOH) is one of the possible feature descriptors in a smile detector. This paper presents an evaluation of the use of Edge Orientation Histograms in a lip image based smile detector. The system built in this paper aims to discriminate lip images depicting a smile (including thin smile and broad smile) from lip images depicting non-smiling expressions. By dividing the lip images into 2 × 4 cells, and using 5° histogram bin size, we achieved 87.8% arithmetic means of accuracies. The experiments show that it is recommended not to use spatial binning that is too small. However, it is recommended to use fine orientation binning. Finally, it is recommended to use all orientation bins as features.