One of the important steps in/towards face recognition is facial feature localization which is difficult. The variety of human faces, expressions, facial hair, race, glasses, poses, and lighting contribute to the complexity of the problem. We systematically varied some of these factors to test the performance of the original active shape model (ASM) and in the next step we modified original ASM that improved accuracy of ASM. Due to these modifications, a new scheme of active shape model for facial feature extraction is proposed in this paper which is named adaptive active shape model. In this scheme, the improvement of the performance of the original ASM focuses on the following three aspects. First, the profile of the original ASM is extended from 1D to 2D. Second, a new face model is constructed in three expressions (natural, smiling, and screaming). Third, in ASM search step, the expression of face is recognized, and the profile model of expression which extracted is used to localize landmark. An extensive experimental investigation is conducted using AR, FERET, JAFFE, YaleB, and Indian face databases. The average improvement of the landmark localization accuracy for proposed method in comparison to the original ASM, under different facial expressions, different illumination conditions, and race variations respectively are 1.85, 0.66 and 7.96.
An Improved Active Shape Model Application on Facial Feature Localization
