A Fast and Rotation Robust Smile Detection Method

Ja-Won Seo and Changick Kim


Smile Detection, maximum coverage, orthogonal projection


In this paper, we propose a fast and rotation robust frontal face smile detection method based on Histograms of Oriented Gradients (HOG). As a feature for classifications, the nature of HOG could be described by the terms of orientation of gradient. Therefore, the majorities of HOG-based smile detections rotate and rescale target face images to make them upright with the expected size before extracting the feature. Unlike those traditional approaches which may require image rotation by the face roll angle (i.e., in-plane rotation), our method directly compensates it in the HOG module and thus restores the original orientations of gradients very fast. Additionally, in order to be irrelevant to the size of faces, we automatically adjust a feature extraction area based on the Inter Pupillary Distance (IPD). As a result, our method outperforms the current state-of-the-art methods in terms of both accuracy and speed. For instance, the experimental results show that the smile detection accuracy of the proposed method is approximately 7.7% higher than Pyramid HOG (PHOG) with only 144 feature dimensions. The computation speed is approximately 2.5 times faster as well.

Important Links:

Go Back