Abstract—One of the most challenge tasks in building a face recognition system is how to represent and extract good quality features from face images. The difficulties come from variations in head poses, illumination conditions, and facial expression. Although many researches have been done, most were carried on under constrained environments. Most researches concentrated on dealing with frontal faces. Processing non-frontal faces encounters more challenge because some features on faces become occluded dramatically. In this paper, we propose two models to extract features from non-frontal faces in the range of 30o to 90o. First, we use the Viola-Jones detection method to identify the pose of face images. Then, we use Active Appearance Model (AAM) to interpret face images. Lastly, the models are trained to know how to fit new images. To improve the efficiency of fitting, we apply a nonlinear parameter update method. Experimental results show that using nonlinear fitting for non-frontal can increase the accuracy of the AAM fitting, compared with some previous methods.

Index Terms—active appearance model, facial feature extraction, non-frontal faces, nonlinear aam fitting, face recognition, expression recognition

Cite: Anh Tuan Do, Thi Duyen Ngo, and The Duy Bui, "Feature Extraction for Non-frontal Faces," Jounal of Automation and Control Engineering, Vol. 4, No. 2, pp. 171-176, April, 2016. doi: 10.12720/joace.4.2.171-176