Enhancing the Detection Rate of Inclined Faces

Extant face detection techniques cannot detect excessively inclined or angled faces, restricting the movement of the subject's facial posture and limiting the scope of face detection applications. Unlike conventional image processing techniques that train classifiers by using rotated frontal face images as positive samples, the researchers of this study employed real-time inclined face images as positive samples and adopted the AdaBoost algorithm for the training procedure. To verify the efficiency of the proposed detection method, the researchers employed three feature extraction methods, namely Haar-like features, histogram of oriented gradients (HOGs), and local binary patterns, to train classifiers from 719 self-developed positive samples and 719 conventional positive samples. Subsequently, a cross-detection experiment was conducted on the sample collections. In addition, the researchers further tested a self-developed video database comprising face videos of 20 subjects. The findings indicate that the proposed detection method outperformed conventional detection methods and improved considerably when coupled with the HOG feature extraction method.