A Mixture of Two Gender Classification Experts

Abstract

This paper presents a novel method for combining the outputs of different gender classification techniques based on facial images. Merging the methods is performed by a committee machine using the Bayesian theorem.We implement and compare several well-known individual classifiers on four different datasets, then we experiment the proposed machine, and show that it significantly improves the accuracy of classification compared to individual classifiers. We also include results that address the effect of scale on the performance of classifiers. I. INTRODUCTIONFacial analysis has been widely investigated in computer vision, including gender, age and expression classification. In particular, gender discrimination is important for several applications; it can improve the performance systems of face verification [1] and face recognition by using separate models for each gender [2], [3], it can help index and retrieve images [4], and it is useful for training interaction systems that behave differently according to the gender of the user.The accuracy of individual gender classification methods can be boosted by merging more than one classifier [5]. When these classifiers use different input features extracted from the face; there is a higher probability that their false classifications on a set of images are disjoint, in which case merge is helpful to minimize the final error of the combined classifier.In this paper we propose a committee machine for merging classification methods based on naive Bayesian theorem, and show, on four different image databases, how this combination improves the performance over the best single constituent classifier by up to more than 4%.The paper is organized as follows; Section II presents an overview of the previous related work. Section III describes the individual classification methods used and Section IV introduces our proposed method for merging these classifiers. Section V explains the experiments we carried and the databases we used, then the results achieved. In the final section, we conclude our work.

Landmarks-SIFT Face Representation for Gender Classification

“…For LFW, we compared with Shan's [11] work and showed comparative results. For the KinFace dataset, we compared with our previous proposal in [17] where we used simple features extracted the face image and combined them at the decision-level achieving 90% accuracy. For FERET, Li et al [3] selected 782 images from the dataset with equal numbers of males and females, and used one quarter for testing and the rest for training.…”

“…To evaluate our proposed method, we used three publicly available face image databases, two of which are benchmark datasets: the FERET [5] and the Labeled Faces in the Wild [6]. The third set: UB KinFace database [7], has not been used before for the purpose of gender classification except for [17], it has been mainly used for the purpose of kinship verification and recognition.…”