Facial Attribute Classification: A Comprehensive Study and a Novel Mid-Level Fusion Classifier

Abdulaziz Alorf, A. Lynn Abbott, Kadir Askin Peker

Abstract: This paper addresses the problem of automatically classifying attributes related to a person's appearance near the face. Example attributes include presence or absence of eyeglasses, separation distance of the eyes (narrow or wide), and information concerning facial hair. We have implemented a 3-step approach in which the face is located and cropped within an image, features are extracted, and 40 separate facial attributes are classified. One goal of the research has been to compare a wide range of techniques for their effectiveness in classifying facial attributes. Another goal has been the development of the new classifier. Our novel system demonstrates better performance than previous state-of-the-art models. We present extensive results to compare our approach with 7 different systems using the CelebA dataset. To our knowledge, this paper presents the most comprehensive study to date concerning facial attributes.